# Transforms¶

## Generic Interfaces¶

### Transform¶

class monai.transforms.Transform[source]

An abstract class of a Transform. A transform is callable that processes data.

It could be stateful and may modify data in place, the implementation should be aware of:

1. thread safety when mutating its own states. When used from a multi-process context, transform’s instance variables are read-only.

2. data content unused by this transform may still be used in the subsequent transforms in a composed transform.

3. storing too much information in data may not scale.

abstract __call__(data)[source]

data is an element which often comes from an iteration over an iterable, such as torch.utils.data.Dataset. This method should return an updated version of data. To simplify the input validations, most of the transforms assume that

• data is a Numpy ndarray, PyTorch Tensor or string

• the data shape can be:

1. string data without shape, LoadImage transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChannel expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirst expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

This method can optionally take additional arguments to help execute transformation operation.

Raises

NotImplementedError – When the subclass does not override this method.

### MapTransform¶

class monai.transforms.MapTransform(keys)[source]

A subclass of monai.transforms.Transform with an assumption that the data input of self.__call__ is a MutableMapping such as dict.

The keys parameter will be used to get and set the actual data item to transform. That is, the callable of this transform should follow the pattern:

def __call__(self, data):
for key in self.keys:
if key in data:
# update output data with some_transform_function(data[key]).
else:
# do nothing or some exceptions handling.
return data
Raises
• ValueError – When keys is an empty iterable.

• TypeError – When keys type is not in Union[Hashable, Iterable[Hashable]].

abstract __call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Returns

An updated dictionary version of data by applying the transform.

### Randomizable¶

class monai.transforms.Randomizable[source]

An interface for handling random state locally, currently based on a class variable R, which is an instance of np.random.RandomState. This is mainly for randomized data augmentation transforms. For example:

class RandShiftIntensity(Randomizable):
def randomize():
self._offset = self.R.uniform(low=0, high=100)
def __call__(self, img):
self.randomize()
return img + self._offset

transform = RandShiftIntensity()
transform.set_random_state(seed=0)
abstract randomize(data)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

set_random_state(seed=None, state=None)[source]

Set the random state locally, to control the randomness, the derived classes should use self.R instead of np.random to introduce random factors.

Parameters
• seed (Optional[int]) – set the random state with an integer seed.

• state (Optional[RandomState]) – set the random state with a np.random.RandomState object.

Raises

TypeError – When state is not an Optional[np.random.RandomState].

Return type

Randomizable

Returns

a Randomizable instance.

### Compose¶

class monai.transforms.Compose(transforms=None)[source]

Compose provides the ability to chain a series of calls together in a sequence. Each transform in the sequence must take a single argument and return a single value, so that the transforms can be called in a chain.

Compose can be used in two ways:

1. With a series of transforms that accept and return a single ndarray / tensor / tensor-like parameter.

2. With a series of transforms that accept and return a dictionary that contains one or more parameters. Such transforms must have pass-through semantics; unused values in the dictionary must be copied to the return dictionary. It is required that the dictionary is copied between input and output of each transform.

If some transform generates a list batch of data in the transform chain, every item in the list is still a dictionary, and all the following transforms will apply to every item of the list, for example:

1. transformA normalizes the intensity of ‘img’ field in the dict data.

2. transformB crops out a list batch of images on ‘img’ and ‘seg’ field. And constructs a list of dict data, other fields are copied:

{                          [{                   {
'img': [1, 2],              'img': [1],         'img': [2],
'seg': [1, 2],              'seg': [1],         'seg': [2],
'extra': 123,    -->        'extra': 123,       'extra': 123,
'shape': 'CHWD'             'shape': 'CHWD'     'shape': 'CHWD'
}                           },                  }]

3. transformC then randomly rotates or flips ‘img’ and ‘seg’ fields of every dictionary item in the list.

The composed transforms will be set the same global random seed if user called set_determinism().

When using the pass-through dictionary operation, you can make use of monai.transforms.adaptors.adaptor to wrap transforms that don’t conform to the requirements. This approach allows you to use transforms from otherwise incompatible libraries with minimal additional work.

Note

In many cases, Compose is not the best way to create pre-processing pipelines. Pre-processing is often not a strictly sequential series of operations, and much of the complexity arises when a not-sequential set of functions must be called as if it were a sequence.

Example: images and labels Images typically require some kind of normalization that labels do not. Both are then typically augmented through the use of random rotations, flips, and deformations. Compose can be used with a series of transforms that take a dictionary that contains ‘image’ and ‘label’ entries. This might require wrapping torchvision transforms before passing them to compose. Alternatively, one can create a class with a __call__ function that calls your pre-processing functions taking into account that not all of them are called on the labels.

__call__(input_)[source]

Call self as a function.

randomize(data=None)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

set_random_state(seed=None, state=None)[source]

Set the random state locally, to control the randomness, the derived classes should use self.R instead of np.random to introduce random factors.

Parameters
• seed (Optional[int]) – set the random state with an integer seed.

• state (Optional[RandomState]) – set the random state with a np.random.RandomState object.

Raises

TypeError – When state is not an Optional[np.random.RandomState].

Return type

Compose

Returns

a Randomizable instance.

## Vanilla Transforms¶

Performs padding to the data, symmetric for all sides or all on one side for each dimension. Uses np.pad so in practice, a mode needs to be provided. See numpy.lib.arraypad.pad for additional details.

Parameters
• spatial_size (Union[Sequence[int], int]) – the spatial size of output data after padding. If its components have non-positive values, the corresponding size of input image will be used (no padding).

• method (Union[Method, str]) – {"symmetric", "end"} Pad image symmetric on every side or only pad at the end sides. Defaults to "symmetric".

• mode (Union[NumpyPadMode, str]) – {"constant", "edge", "linear_ramp", "maximum", "mean", "median", "minimum", "reflect", "symmetric", "wrap", "empty"} One of the listed string values or a user supplied function. Defaults to "constant". See also: https://numpy.org/doc/1.18/reference/generated/numpy.pad.html

__call__(img, mode=None)[source]
Parameters
• img (ndarray) – data to be transformed, assuming img is channel-first and padding doesn’t apply to the channel dim.

• mode (Union[NumpyPadMode, str, None]) – {"constant", "edge", "linear_ramp", "maximum", "mean", "median", "minimum", "reflect", "symmetric", "wrap", "empty"} One of the listed string values or a user supplied function. Defaults to self.mode. See also: https://numpy.org/doc/1.18/reference/generated/numpy.pad.html

Return type

ndarray

Parameters
• spatial_border (Union[Sequence[int], int]) –

specified size for every spatial border. it can be 3 shapes:

• single int number, pad all the borders with the same size.

• length equals the length of image shape, pad every spatial dimension separately. for example, image shape(CHW) is [1, 4, 4], spatial_border is [2, 1], pad every border of H dim with 2, pad every border of W dim with 1, result shape is [1, 8, 6].

• length equals 2 x (length of image shape), pad every border of every dimension separately. for example, image shape(CHW) is [1, 4, 4], spatial_border is [1, 2, 3, 4], pad top of H dim with 1, pad bottom of H dim with 2, pad left of W dim with 3, pad right of W dim with 4. the result shape is [1, 7, 11].

• mode (Union[NumpyPadMode, str]) – {"constant", "edge", "linear_ramp", "maximum", "mean", "median", "minimum", "reflect", "symmetric", "wrap", "empty"} One of the listed string values or a user supplied function. Defaults to "constant". See also: https://numpy.org/doc/1.18/reference/generated/numpy.pad.html

__call__(img, mode=None)[source]
Parameters
• img (ndarray) – data to be transformed, assuming img is channel-first and padding doesn’t apply to the channel dim.

• mode (Union[NumpyPadMode, str, None]) – {"constant", "edge", "linear_ramp", "maximum", "mean", "median", "minimum", "reflect", "symmetric", "wrap", "empty"} One of the listed string values or a user supplied function. Defaults to self.mode. See also: https://numpy.org/doc/1.18/reference/generated/numpy.pad.html

Raises
• ValueError – When self.spatial_border contains a nonnegative int.

• ValueError – When self.spatial_border length is not one of [1, len(spatial_shape), 2*len(spatial_shape)].

Return type

ndarray

Pad the input data, so that the spatial sizes are divisible by k.

Parameters
• k (Union[Sequence[int], int]) – the target k for each spatial dimension. if k is negative or 0, the original size is preserved. if k is an int, the same k be applied to all the input spatial dimensions.

• mode (Union[NumpyPadMode, str]) – {"constant", "edge", "linear_ramp", "maximum", "mean", "median", "minimum", "reflect", "symmetric", "wrap", "empty"} One of the listed string values or a user supplied function. Defaults to "constant". See also: https://numpy.org/doc/1.18/reference/generated/numpy.pad.html

__call__(img, mode=None)[source]
Parameters
• img (ndarray) – data to be transformed, assuming img is channel-first and padding doesn’t apply to the channel dim.

• mode (Union[NumpyPadMode, str, None]) – {"constant", "edge", "linear_ramp", "maximum", "mean", "median", "minimum", "reflect", "symmetric", "wrap", "empty"} One of the listed string values or a user supplied function. Defaults to self.mode. See also: https://numpy.org/doc/1.18/reference/generated/numpy.pad.html

Return type

ndarray

#### SpatialCrop¶

class monai.transforms.SpatialCrop(roi_center=None, roi_size=None, roi_start=None, roi_end=None)[source]

General purpose cropper to produce sub-volume region of interest (ROI). It can support to crop ND spatial (channel-first) data. Either a spatial center and size must be provided, or alternatively, if center and size are not provided, the start and end coordinates of the ROI must be provided.

Parameters
• roi_center (Optional[Sequence[int]]) – voxel coordinates for center of the crop ROI.

• roi_size (Optional[Sequence[int]]) – size of the crop ROI.

• roi_start (Optional[Sequence[int]]) – voxel coordinates for start of the crop ROI.

• roi_end (Optional[Sequence[int]]) – voxel coordinates for end of the crop ROI.

__call__(img)[source]

Apply the transform to img, assuming img is channel-first and slicing doesn’t apply to the channel dim.

Return type

ndarray

#### CenterSpatialCrop¶

class monai.transforms.CenterSpatialCrop(roi_size)[source]

Crop at the center of image with specified ROI size.

Parameters

roi_size (Union[Sequence[int], int]) – the spatial size of the crop region e.g. [224,224,128] If its components have non-positive values, the corresponding size of input image will be used.

__call__(img)[source]

Apply the transform to img, assuming img is channel-first and slicing doesn’t apply to the channel dim.

Return type

ndarray

#### RandSpatialCrop¶

class monai.transforms.RandSpatialCrop(roi_size, random_center=True, random_size=True)[source]

Crop image with random size or specific size ROI. It can crop at a random position as center or at the image center. And allows to set the minimum size to limit the randomly generated ROI.

Parameters
• roi_size (Union[Sequence[int], int]) – if random_size is True, it specifies the minimum crop region. if random_size is False, it specifies the expected ROI size to crop. e.g. [224, 224, 128] If its components have non-positive values, the corresponding size of input image will be used.

• random_center (bool) – crop at random position as center or the image center.

• random_size (bool) – crop with random size or specific size ROI. The actual size is sampled from randint(roi_size, img_size).

__call__(img)[source]

Apply the transform to img, assuming img is channel-first and slicing doesn’t apply to the channel dim.

Return type

ndarray

randomize(img_size)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

#### RandSpatialCropSamples¶

class monai.transforms.RandSpatialCropSamples(roi_size, num_samples, random_center=True, random_size=True)[source]

Crop image with random size or specific size ROI to generate a list of N samples. It can crop at a random position as center or at the image center. And allows to set the minimum size to limit the randomly generated ROI. It will return a list of cropped images.

Parameters
• roi_size (Union[Sequence[int], int]) – if random_size is True, the spatial size of the minimum crop region. if random_size is False, specify the expected ROI size to crop. e.g. [224, 224, 128]

• num_samples (int) – number of samples (crop regions) to take in the returned list.

• random_center (bool) – crop at random position as center or the image center.

• random_size (bool) – crop with random size or specific size ROI. The actual size is sampled from randint(roi_size, img_size).

Raises

ValueError – When num_samples is nonpositive.

__call__(img)[source]

Apply the transform to img, assuming img is channel-first and cropping doesn’t change the channel dim.

Return type

List[ndarray]

randomize(data=None)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

set_random_state(seed=None, state=None)[source]

Set the random state locally, to control the randomness, the derived classes should use self.R instead of np.random to introduce random factors.

Parameters
• seed (Optional[int]) – set the random state with an integer seed.

• state (Optional[RandomState]) – set the random state with a np.random.RandomState object.

Raises

TypeError – When state is not an Optional[np.random.RandomState].

Return type

Randomizable

Returns

a Randomizable instance.

#### CropForeground¶

class monai.transforms.CropForeground(select_fn=<function CropForeground.<lambda>>, channel_indices=None, margin=0, return_coords=False)[source]

Crop an image using a bounding box. The bounding box is generated by selecting foreground using select_fn at channels channel_indices. margin is added in each spatial dimension of the bounding box. The typical usage is to help training and evaluation if the valid part is small in the whole medical image. Users can define arbitrary function to select expected foreground from the whole image or specified channels. And it can also add margin to every dim of the bounding box of foreground object. For example:

image = np.array(
[[[0, 0, 0, 0, 0],
[0, 1, 2, 1, 0],
[0, 1, 3, 2, 0],
[0, 1, 2, 1, 0],
[0, 0, 0, 0, 0]]])  # 1x5x5, single channel 5x5 image
cropper = CropForeground(select_fn=lambda x: x > 1, margin=0)
print(cropper(image))
[[[2, 1],
[3, 2],
[2, 1]]]
Parameters
• select_fn (Callable) – function to select expected foreground, default is to select values > 0.

• channel_indices (Union[Iterable[int], int, None]) – if defined, select foreground only on the specified channels of image. if None, select foreground on the whole image.

• margin (Union[Sequence[int], int]) – add margin value to spatial dims of the bounding box, if only 1 value provided, use it for all dims.

• return_coords (bool) – whether return the coordinates of spatial bounding box for foreground.

__call__(img)[source]

Apply the transform to img, assuming img is channel-first and slicing doesn’t change the channel dim.

#### RandWeightedCrop¶

class monai.transforms.RandWeightedCrop(spatial_size, num_samples=1, weight_map=None)[source]

Samples a list of num_samples image patches according to the provided weight_map.

Parameters
• spatial_size (Union[Sequence[int], int]) – the spatial size of the image patch e.g. [224, 224, 128]. If its components have non-positive values, the corresponding size of img will be used.

• num_samples (int) – number of samples (image patches) to take in the returned list.

• weight_map (Optional[ndarray]) – weight map used to generate patch samples. The weights must be non-negative. Each element denotes a sampling weight of the spatial location. 0 indicates no sampling. It should be a single-channel array in shape, for example, (1, spatial_dim_0, spatial_dim_1, …).

__call__(img, weight_map=None)[source]
Parameters
• img (ndarray) – input image to sample patches from. assuming img is a channel-first array.

• weight_map (Optional[ndarray]) – weight map used to generate patch samples. The weights must be non-negative. Each element denotes a sampling weight of the spatial location. 0 indicates no sampling. It should be a single-channel array in shape, for example, (1, spatial_dim_0, spatial_dim_1, …)

Return type

List[ndarray]

Returns

A list of image patches

randomize(weight_map)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

#### RandCropByPosNegLabel¶

class monai.transforms.RandCropByPosNegLabel(spatial_size, label=None, pos=1.0, neg=1.0, num_samples=1, image=None, image_threshold=0.0, fg_indices=None, bg_indices=None)[source]

Crop random fixed sized regions with the center being a foreground or background voxel based on the Pos Neg Ratio. And will return a list of arrays for all the cropped images. For example, crop two (3 x 3) arrays from (5 x 5) array with pos/neg=1:

[[[0, 0, 0, 0, 0],
[0, 1, 2, 1, 0],            [[0, 1, 2],     [[2, 1, 0],
[0, 1, 3, 0, 0],     -->     [0, 1, 3],      [3, 0, 0],
[0, 0, 0, 0, 0],             [0, 0, 0]]      [0, 0, 0]]
[0, 0, 0, 0, 0]]]
Parameters
• spatial_size (Union[Sequence[int], int]) – the spatial size of the crop region e.g. [224, 224, 128]. If its components have non-positive values, the corresponding size of label will be used.

• label (Optional[ndarray]) – the label image that is used for finding foreground/background, if None, must set at self.__call__. Non-zero indicates foreground, zero indicates background.

• pos (float) – used with neg together to calculate the ratio pos / (pos + neg) for the probability to pick a foreground voxel as a center rather than a background voxel.

• neg (float) – used with pos together to calculate the ratio pos / (pos + neg) for the probability to pick a foreground voxel as a center rather than a background voxel.

• num_samples (int) – number of samples (crop regions) to take in each list.

• image (Optional[ndarray]) – optional image data to help select valid area, can be same as img or another image array. if not None, use label == 0 & image > image_threshold to select the negative sample (background) center. So the crop center will only come from the valid image areas.

• image_threshold (float) – if enabled image, use image > image_threshold to determine the valid image content areas.

• fg_indices (Optional[ndarray]) – if provided pre-computed foreground indices of label, will ignore above image and image_threshold, and randomly select crop centers based on them, need to provide fg_indices and bg_indices together, expect to be 1 dim array of spatial indices after flattening. a typical usage is to call FgBgToIndices transform first and cache the results.

• bg_indices (Optional[ndarray]) – if provided pre-computed background indices of label, will ignore above image and image_threshold, and randomly select crop centers based on them, need to provide fg_indices and bg_indices together, expect to be 1 dim array of spatial indices after flattening. a typical usage is to call FgBgToIndices transform first and cache the results.

Raises
• ValueError – When pos or neg are negative.

• ValueError – When pos=0 and neg=0. Incompatible values.

__call__(img, label=None, image=None, fg_indices=None, bg_indices=None)[source]
Parameters
• img (ndarray) – input data to crop samples from based on the pos/neg ratio of label and image. Assumes img is a channel-first array.

• label (Optional[ndarray]) – the label image that is used for finding foreground/background, if None, use self.label.

• image (Optional[ndarray]) – optional image data to help select valid area, can be same as img or another image array. use label == 0 & image > image_threshold to select the negative sample(background) center. so the crop center will only exist on valid image area. if None, use self.image.

• fg_indices (Optional[ndarray]) – foreground indices to randomly select crop centers, need to provide fg_indices and bg_indices together.

• bg_indices (Optional[ndarray]) – background indices to randomly select crop centers, need to provide fg_indices and bg_indices together.

Return type

List[ndarray]

randomize(label, fg_indices=None, bg_indices=None, image=None)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

Resize an image to a target spatial size by either centrally cropping the image or padding it evenly with a user-specified mode. When the dimension is smaller than the target size, do symmetric padding along that dim. When the dimension is larger than the target size, do central cropping along that dim.

Parameters
• spatial_size (Union[Sequence[int], int]) – the spatial size of output data after padding or crop. If has non-positive values, the corresponding size of input image will be used (no padding).

• mode (Union[NumpyPadMode, str]) – {"constant", "edge", "linear_ramp", "maximum", "mean", "median", "minimum", "reflect", "symmetric", "wrap", "empty"} One of the listed string values or a user supplied function for padding. Defaults to "constant". See also: https://numpy.org/doc/1.18/reference/generated/numpy.pad.html

__call__(img, mode=None)[source]
Parameters
• img (ndarray) – data to pad or crop, assuming img is channel-first and padding or cropping doesn’t apply to the channel dim.

• mode (Union[NumpyPadMode, str, None]) – {"constant", "edge", "linear_ramp", "maximum", "mean", "median", "minimum", "reflect", "symmetric", "wrap", "empty"} One of the listed string values or a user supplied function for padding. If None, defaults to the mode in construction. See also: https://numpy.org/doc/1.18/reference/generated/numpy.pad.html

Return type

ndarray

#### BoundingRect¶

class monai.transforms.BoundingRect(select_fn=<function BoundingRect.<lambda>>)[source]

Compute coordinates of axis-aligned bounding rectangles from input image img. The output format of the coordinates is (shape is [channel, 2 * spatial dims]):

[[1st_spatial_dim_start, 1st_spatial_dim_end,

2nd_spatial_dim_start, 2nd_spatial_dim_end, …, Nth_spatial_dim_start, Nth_spatial_dim_end],

[1st_spatial_dim_start, 1st_spatial_dim_end, 2nd_spatial_dim_start, 2nd_spatial_dim_end, …, Nth_spatial_dim_start, Nth_spatial_dim_end]]

The bounding boxes edges are aligned with the input image edges. This function returns [-1, -1, …] if there’s no positive intensity.

Parameters

select_fn (Callable) – function to select expected foreground, default is to select values > 0.

__call__(img)[source]
Return type

ndarray

### Intensity¶

#### RandGaussianNoise¶

class monai.transforms.RandGaussianNoise(prob=0.1, mean=0.0, std=0.1)[source]

Parameters
• prob (float) – Probability to add Gaussian noise.

• mean (Union[Sequence[float], float]) – Mean or “centre” of the distribution.

• std (float) – Standard deviation (spread) of distribution.

__call__(img)[source]

Apply the transform to img.

Return type

Union[Tensor, ndarray]

randomize(im_shape)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

#### ShiftIntensity¶

class monai.transforms.ShiftIntensity(offset)[source]

Shift intensity uniformly for the entire image with specified offset.

Parameters

offset (float) – offset value to shift the intensity of image.

__call__(img)[source]

Apply the transform to img.

Return type

ndarray

#### RandShiftIntensity¶

class monai.transforms.RandShiftIntensity(offsets, prob=0.1)[source]

Randomly shift intensity with randomly picked offset.

Parameters
• offsets (Union[Tuple[float, float], float]) – offset range to randomly shift. if single number, offset value is picked from (-offsets, offsets).

• prob (float) – probability of shift.

__call__(img)[source]

Apply the transform to img.

Return type

ndarray

randomize(data=None)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

#### ScaleIntensity¶

class monai.transforms.ScaleIntensity(minv=0.0, maxv=1.0, factor=None)[source]

Scale the intensity of input image to the given value range (minv, maxv). If minv and maxv not provided, use factor to scale image by v = v * (1 + factor).

Parameters
• minv (Optional[float]) – minimum value of output data.

• maxv (Optional[float]) – maximum value of output data.

• factor (Optional[float]) – factor scale by v = v * (1 + factor).

__call__(img)[source]

Apply the transform to img.

Raises

ValueError – When self.minv=None or self.maxv=None and self.factor=None. Incompatible values.

Return type

ndarray

#### RandScaleIntensity¶

class monai.transforms.RandScaleIntensity(factors, prob=0.1)[source]

Randomly scale the intensity of input image by v = v * (1 + factor) where the factor is randomly picked from (-factors[0], factors[0]).

Parameters
• factors (Union[Tuple[float, float], float]) – factor range to randomly scale by v = v * (1 + factor). if single number, factor value is picked from (-factors, factors).

• prob (float) – probability of scale.

__call__(img)[source]

Apply the transform to img.

Return type

ndarray

randomize(data=None)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

#### NormalizeIntensity¶

class monai.transforms.NormalizeIntensity(subtrahend=None, divisor=None, nonzero=False, channel_wise=False, dtype=<class 'numpy.float32'>)[source]

Normalize input based on provided args, using calculated mean and std if not provided. This transform can normalize only non-zero values or entire image, and can also calculate mean and std on each channel separately. When channel_wise is True, the first dimension of subtrahend and divisor should be the number of image channels if they are not None.

Parameters
• subtrahend (Optional[Sequence]) – the amount to subtract by (usually the mean).

• divisor (Optional[Sequence]) – the amount to divide by (usually the standard deviation).

• nonzero (bool) – whether only normalize non-zero values.

• channel_wise (bool) – if using calculated mean and std, calculate on each channel separately or calculate on the entire image directly.

• dtype (dtype) – output data type, defaut to float32.

__call__(img)[source]

Apply the transform to img, assuming img is a channel-first array if self.channel_wise is True,

Return type

ndarray

#### ThresholdIntensity¶

class monai.transforms.ThresholdIntensity(threshold, above=True, cval=0.0)[source]

Filter the intensity values of whole image to below threshold or above threshold. And fill the remaining parts of the image to the cval value.

Parameters
• threshold (float) – the threshold to filter intensity values.

• above (bool) – filter values above the threshold or below the threshold, default is True.

• cval (float) – value to fill the remaining parts of the image, default is 0.

__call__(img)[source]

Apply the transform to img.

Return type

ndarray

#### ScaleIntensityRange¶

class monai.transforms.ScaleIntensityRange(a_min, a_max, b_min, b_max, clip=False)[source]

Apply specific intensity scaling to the whole numpy array. Scaling from [a_min, a_max] to [b_min, b_max] with clip option.

Parameters
• a_min (float) – intensity original range min.

• a_max (float) – intensity original range max.

• b_min (float) – intensity target range min.

• b_max (float) – intensity target range max.

• clip (bool) – whether to perform clip after scaling.

__call__(img)[source]

Apply the transform to img.

Return type

ndarray

#### ScaleIntensityRangePercentiles¶

class monai.transforms.ScaleIntensityRangePercentiles(lower, upper, b_min, b_max, clip=False, relative=False)[source]

Apply range scaling to a numpy array based on the intensity distribution of the input.

By default this transform will scale from [lower_intensity_percentile, upper_intensity_percentile] to [b_min, b_max], where {lower,upper}_intensity_percentile are the intensity values at the corresponding percentiles of img.

The relative parameter can also be set to scale from [lower_intensity_percentile, upper_intensity_percentile] to the lower and upper percentiles of the output range [b_min, b_max]

For example:

image = np.array(
[[[1, 2, 3, 4, 5],
[1, 2, 3, 4, 5],
[1, 2, 3, 4, 5],
[1, 2, 3, 4, 5],
[1, 2, 3, 4, 5],
[1, 2, 3, 4, 5]]])

# Scale from lower and upper image intensity percentiles
# to output range [b_min, b_max]
scaler = ScaleIntensityRangePercentiles(10, 90, 0, 200, False, False)
print(scaler(image))
[[[0., 50., 100., 150., 200.],
[0., 50., 100., 150., 200.],
[0., 50., 100., 150., 200.],
[0., 50., 100., 150., 200.],
[0., 50., 100., 150., 200.],
[0., 50., 100., 150., 200.]]]

# Scale from lower and upper image intensity percentiles
# to lower and upper percentiles of the output range [b_min, b_max]
rel_scaler = ScaleIntensityRangePercentiles(10, 90, 0, 200, False, True)
print(rel_scaler(image))
[[[20., 60., 100., 140., 180.],
[20., 60., 100., 140., 180.],
[20., 60., 100., 140., 180.],
[20., 60., 100., 140., 180.],
[20., 60., 100., 140., 180.],
[20., 60., 100., 140., 180.]]]
Parameters
• lower (float) – lower intensity percentile.

• upper (float) – upper intensity percentile.

• b_min (float) – intensity target range min.

• b_max (float) – intensity target range max.

• clip (bool) – whether to perform clip after scaling.

• relative (bool) – whether to scale to the corresponding percentiles of [b_min, b_max].

__call__(img)[source]

Apply the transform to img.

Return type

ndarray

Changes image intensity by gamma. Each pixel/voxel intensity is updated as:

x = ((x - min) / intensity_range) ^ gamma * intensity_range + min
Parameters

gamma (float) – gamma value to adjust the contrast as function.

__call__(img)[source]

Apply the transform to img.

Return type

ndarray

Randomly changes image intensity by gamma. Each pixel/voxel intensity is updated as:

x = ((x - min) / intensity_range) ^ gamma * intensity_range + min
Parameters
• prob (float) – Probability of adjustment.

• gamma (Union[Sequence[float], float]) – Range of gamma values. If single number, value is picked from (0.5, gamma), default is (0.5, 4.5).

__call__(img)[source]

Apply the transform to img.

Return type

ndarray

randomize(data=None)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

Mask the intensity values of input image with the specified mask data. Mask data must have the same spatial size as the input image, and all the intensity values of input image corresponding to 0 in the mask data will be set to 0, others will keep the original value.

Parameters

mask_data (ndarray) – if mask data is single channel, apply to evey channel of input image. if multiple channels, the channel number must match input data. mask_data will be converted to bool values by mask_data > 0 before applying transform to input image.

Parameters

mask_data (Optional[ndarray]) – if mask data is single channel, apply to evey channel of input image. if multiple channels, the channel number must match input data. mask_data will be converted to bool values by mask_data > 0 before applying transform to input image.

Raises

ValueError – When mask_data and img channels differ and mask_data is not single channel.

Return type

ndarray

#### SavitzkyGolaySmooth¶

class monai.transforms.SavitzkyGolaySmooth(window_length, order, axis=1, mode='zeros')[source]

Smooth the input data along the given axis using a Savitzky-Golay filter.

Parameters
• window_length (int) – Length of the filter window, must be a positive odd integer.

• order (int) – Order of the polynomial to fit to each window, must be less than window_length.

• axis (int) – Optional axis along which to apply the filter kernel. Default 1 (first spatial dimension).

• mode (str) – Optional padding mode, passed to convolution class. 'zeros', 'reflect', 'replicate' or 'circular'. Default: 'zeros'. See torch.nn.Conv1d() for more information.

__call__(img)[source]
Parameters

img (ndarray) – numpy.ndarray containing input data. Must be real and in shape [channels, spatial1, spatial2, …].

Return type

ndarray

Returns

np.ndarray containing smoothed result.

#### GaussianSmooth¶

class monai.transforms.GaussianSmooth(sigma=1.0, approx='erf')[source]

Apply Gaussian smooth to the input data based on specified sigma parameter. A default value sigma=1.0 is provided for reference.

Parameters
• sigma (Union[Sequence[float], float]) – if a list of values, must match the count of spatial dimensions of input data, and apply every value in the list to 1 spatial dimension. if only 1 value provided, use it for all spatial dimensions.

• approx (str) – discrete Gaussian kernel type, available options are “erf”, “sampled”, and “scalespace”. see also monai.networks.layers.GaussianFilter().

__call__(img)[source]

data is an element which often comes from an iteration over an iterable, such as torch.utils.data.Dataset. This method should return an updated version of data. To simplify the input validations, most of the transforms assume that

• data is a Numpy ndarray, PyTorch Tensor or string

• the data shape can be:

1. string data without shape, LoadImage transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChannel expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirst expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

This method can optionally take additional arguments to help execute transformation operation.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

ndarray

#### RandGaussianSmooth¶

class monai.transforms.RandGaussianSmooth(sigma_x=(0.25, 1.5), sigma_y=(0.25, 1.5), sigma_z=(0.25, 1.5), prob=0.1, approx='erf')[source]

Apply Gaussian smooth to the input data based on randomly selected sigma parameters.

Parameters
• sigma_x (Tuple[float, float]) – randomly select sigma value for the first spatial dimension.

• sigma_y (Tuple[float, float]) – randomly select sigma value for the second spatial dimension if have.

• sigma_z (Tuple[float, float]) – randomly select sigma value for the third spatial dimension if have.

• prob (float) – probability of Gaussian smooth.

• approx (str) – discrete Gaussian kernel type, available options are “erf”, “sampled”, and “scalespace”. see also monai.networks.layers.GaussianFilter().

__call__(img)[source]

Call self as a function.

Return type

ndarray

randomize(data=None)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

#### GaussianSharpen¶

class monai.transforms.GaussianSharpen(sigma1=3.0, sigma2=1.0, alpha=30.0, approx='erf')[source]

Sharpen images using the Gaussian Blur filter. Referring to: http://scipy-lectures.org/advanced/image_processing/auto_examples/plot_sharpen.html. The algorithm is shown as below

blurred_f = gaussian_filter(img, sigma1)
filter_blurred_f = gaussian_filter(blurred_f, sigma2)
img = blurred_f + alpha * (blurred_f - filter_blurred_f)

A set of default values sigma1=3.0, sigma2=1.0 and alpha=30.0 is provide for reference.

Parameters
• sigma1 (Union[Sequence[float], float]) – sigma parameter for the first gaussian kernel. if a list of values, must match the count of spatial dimensions of input data, and apply every value in the list to 1 spatial dimension. if only 1 value provided, use it for all spatial dimensions.

• sigma2 (Union[Sequence[float], float]) – sigma parameter for the second gaussian kernel. if a list of values, must match the count of spatial dimensions of input data, and apply every value in the list to 1 spatial dimension. if only 1 value provided, use it for all spatial dimensions.

• alpha (float) – weight parameter to compute the final result.

• approx (str) – discrete Gaussian kernel type, available options are “erf”, “sampled”, and “scalespace”. see also monai.networks.layers.GaussianFilter().

__call__(img)[source]

data is an element which often comes from an iteration over an iterable, such as torch.utils.data.Dataset. This method should return an updated version of data. To simplify the input validations, most of the transforms assume that

• data is a Numpy ndarray, PyTorch Tensor or string

• the data shape can be:

1. string data without shape, LoadImage transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChannel expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirst expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

This method can optionally take additional arguments to help execute transformation operation.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

ndarray

#### RandGaussianSharpen¶

class monai.transforms.RandGaussianSharpen(sigma1_x=(0.5, 1.0), sigma1_y=(0.5, 1.0), sigma1_z=(0.5, 1.0), sigma2_x=0.5, sigma2_y=0.5, sigma2_z=0.5, alpha=(10.0, 30.0), approx='erf', prob=0.1)[source]

Sharpen images using the Gaussian Blur filter based on randomly selected sigma1, sigma2 and alpha. The algorithm is monai.transforms.GaussianSharpen.

Parameters
• sigma1_x (Tuple[float, float]) – randomly select sigma value for the first spatial dimension of first gaussian kernel.

• sigma1_y (Tuple[float, float]) – randomly select sigma value for the second spatial dimension(if have) of first gaussian kernel.

• sigma1_z (Tuple[float, float]) – randomly select sigma value for the third spatial dimension(if have) of first gaussian kernel.

• sigma2_x (Union[Tuple[float, float], float]) – randomly select sigma value for the first spatial dimension of second gaussian kernel. if only 1 value X provided, it must be smaller than sigma1_x and randomly select from [X, sigma1_x].

• sigma2_y (Union[Tuple[float, float], float]) – randomly select sigma value for the second spatial dimension(if have) of second gaussian kernel. if only 1 value Y provided, it must be smaller than sigma1_y and randomly select from [Y, sigma1_y].

• sigma2_z (Union[Tuple[float, float], float]) – randomly select sigma value for the third spatial dimension(if have) of second gaussian kernel. if only 1 value Z provided, it must be smaller than sigma1_z and randomly select from [Z, sigma1_z].

• alpha (Tuple[float, float]) – randomly select weight parameter to compute the final result.

• approx (str) – discrete Gaussian kernel type, available options are “erf”, “sampled”, and “scalespace”. see also monai.networks.layers.GaussianFilter().

• prob (float) – probability of Gaussian sharpen.

__call__(img)[source]

Call self as a function.

Return type

ndarray

randomize(data=None)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

#### RandHistogramShift¶

class monai.transforms.RandHistogramShift(num_control_points=10, prob=0.1)[source]

Apply random nonlinear transform to the image’s intensity histogram.

Parameters
• num_control_points (Union[Tuple[int, int], int]) – number of control points governing the nonlinear intensity mapping. a smaller number of control points allows for larger intensity shifts. if two values provided, number of control points selecting from range (min_value, max_value).

• prob (float) – probability of histogram shift.

__call__(img)[source]

Call self as a function.

Return type

ndarray

randomize(data=None)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

#### DetectEnvelope¶

class monai.transforms.DetectEnvelope(axis=1, n=None)[source]

Find the envelope of the input data along the requested axis using a Hilbert transform. Requires PyTorch 1.7.0+ and the PyTorch FFT module (which is not included in NVIDIA PyTorch Release 20.10).

Parameters
• axis (int) – Axis along which to detect the envelope. Default 1, i.e. the first spatial dimension.

• N – FFT size. Default img.shape[axis]. Input will be zero-padded or truncated to this size along dimension

• axis.

__call__(img)[source]
Parameters

img (ndarray) – numpy.ndarray containing input data. Must be real and in shape [channels, spatial1, spatial2, …].

Return type

ndarray

Returns

np.ndarray containing envelope of data in img along the specified axis.

### IO¶

Parameters

• image_only (bool) – if True return only the image volume, otherwise return image data array and header dict.

• dtype (dtype) – if not None convert the loaded image to this data type.

Note

The transform returns image data array if image_only is True, or a tuple of two elements containing the data array, and the meta data in a dict format otherwise.

Parameters
• filename (Union[Sequence[str], str]) – path file or file-like object or a list of files. will save the filename to meta_data with key filename_or_obj. if provided a list of files, use the filename of first file.

Parameters

Return type

### Post-processing¶

#### Activations¶

class monai.transforms.Activations(sigmoid=False, softmax=False, other=None)[source]

Add activation operations to the model output, typically Sigmoid or Softmax.

Parameters
• sigmoid (bool) – whether to execute sigmoid function on model output before transform. Defaults to False.

• softmax (bool) – whether to execute softmax function on model output before transform. Defaults to False.

• other (Optional[Callable]) – callable function to execute other activation layers, for example: other = lambda x: torch.tanh(x). Defaults to None.

Raises

TypeError – When other is not an Optional[Callable].

__call__(img, sigmoid=None, softmax=None, other=None)[source]
Parameters
• sigmoid (Optional[bool]) – whether to execute sigmoid function on model output before transform. Defaults to self.sigmoid.

• softmax (Optional[bool]) – whether to execute softmax function on model output before transform. Defaults to self.softmax.

• other (Optional[Callable]) – callable function to execute other activation layers, for example: other = lambda x: torch.tanh(x). Defaults to self.other.

Raises
• ValueError – When sigmoid=True and softmax=True. Incompatible values.

• TypeError – When other is not an Optional[Callable].

• ValueError – When self.other=None and other=None. Incompatible values.

Return type

Tensor

#### AsDiscrete¶

class monai.transforms.AsDiscrete(argmax=False, to_onehot=False, n_classes=None, threshold_values=False, logit_thresh=0.5)[source]

Execute after model forward to transform model output to discrete values. It can complete below operations:

• execute argmax for input logits values.

• threshold input value to 0.0 or 1.0.

• convert input value to One-Hot format

Parameters
• argmax (bool) – whether to execute argmax function on input data before transform. Defaults to False.

• to_onehot (bool) – whether to convert input data into the one-hot format. Defaults to False.

• n_classes (Optional[int]) – the number of classes to convert to One-Hot format. Defaults to None.

• threshold_values (bool) – whether threshold the float value to int number 0 or 1. Defaults to False.

• logit_thresh (float) – the threshold value for thresholding operation.. Defaults to 0.5.

__call__(img, argmax=None, to_onehot=None, n_classes=None, threshold_values=None, logit_thresh=None)[source]
Parameters
• argmax (Optional[bool]) – whether to execute argmax function on input data before transform. Defaults to self.argmax.

• to_onehot (Optional[bool]) – whether to convert input data into the one-hot format. Defaults to self.to_onehot.

• n_classes (Optional[int]) – the number of classes to convert to One-Hot format. Defaults to self.n_classes.

• threshold_values (Optional[bool]) – whether threshold the float value to int number 0 or 1. Defaults to self.threshold_values.

• logit_thresh (Optional[float]) – the threshold value for thresholding operation.. Defaults to self.logit_thresh.

Return type

Tensor

#### KeepLargestConnectedComponent¶

class monai.transforms.KeepLargestConnectedComponent(applied_labels, independent=True, connectivity=None)[source]

Keeps only the largest connected component in the image. This transform can be used as a post-processing step to clean up over-segment areas in model output.

The input is assumed to be a PyTorch Tensor:
1. With shape (batch_size, 1, spatial_dim1[, spatial_dim2, …]) and the values correspond to expected labels.

2. With shape (batch_size, C, spatial_dim1[, spatial_dim2, …]) and the values should be 0, 1 on each labels.

Note

For single channel data, 0 will be treated as background and the over-segment pixels will be set to 0. For one-hot data, the over-segment pixels will be set to 0 in its channel.

For example: Use KeepLargestConnectedComponent with applied_labels=[1], connectivity=1:

[1, 0, 0]         [0, 0, 0]
[0, 1, 1]    =>   [0, 1 ,1]
[0, 1, 1]         [0, 1, 1]

Use KeepLargestConnectedComponent with applied_labels[1, 2], independent=False, connectivity=1:

[0, 0, 1, 0 ,0]           [0, 0, 1, 0 ,0]
[0, 2, 1, 1 ,1]           [0, 2, 1, 1 ,1]
[1, 2, 1, 0 ,0]    =>     [1, 2, 1, 0 ,0]
[1, 2, 0, 1 ,0]           [1, 2, 0, 0 ,0]
[2, 2, 0, 0 ,2]           [2, 2, 0, 0 ,0]

Use KeepLargestConnectedComponent with applied_labels[1, 2], independent=True, connectivity=1:

[0, 0, 1, 0 ,0]           [0, 0, 1, 0 ,0]
[0, 2, 1, 1 ,1]           [0, 2, 1, 1 ,1]
[1, 2, 1, 0 ,0]    =>     [0, 2, 1, 0 ,0]
[1, 2, 0, 1 ,0]           [0, 2, 0, 0 ,0]
[2, 2, 0, 0 ,2]           [2, 2, 0, 0 ,0]

Use KeepLargestConnectedComponent with applied_labels[1, 2], independent=False, connectivity=2:

[0, 0, 1, 0 ,0]           [0, 0, 1, 0 ,0]
[0, 2, 1, 1 ,1]           [0, 2, 1, 1 ,1]
[1, 2, 1, 0 ,0]    =>     [1, 2, 1, 0 ,0]
[1, 2, 0, 1 ,0]           [1, 2, 0, 1 ,0]
[2, 2, 0, 0 ,2]           [2, 2, 0, 0 ,2]
Parameters
• applied_labels (Union[Sequence[int], int]) – Labels for applying the connected component on. If only one channel. The pixel whose value is not in this list will remain unchanged. If the data is in one-hot format, this is used to determine what channels to apply.

• independent (bool) – consider several labels as a whole or independent, default is True. Example use case would be segment label 1 is liver and label 2 is liver tumor, in that case you want this “independent” to be specified as False.

• connectivity (Optional[int]) – Maximum number of orthogonal hops to consider a pixel/voxel as a neighbor. Accepted values are ranging from 1 to input.ndim. If None, a full connectivity of input.ndim is used.

__call__(img)[source]
Parameters

img (Tensor) – shape must be (batch_size, C, spatial_dim1[, spatial_dim2, …]).

Return type

Tensor

Returns

A PyTorch Tensor with shape (batch_size, C, spatial_dim1[, spatial_dim2, …]).

#### LabelToContour¶

class monai.transforms.LabelToContour(kernel_type='Laplace')[source]

Return the contour of binary input images that only compose of 0 and 1, with Laplace kernel set as default for edge detection. Typical usage is to plot the edge of label or segmentation output.

Parameters

kernel_type (str) – the method applied to do edge detection, default is “Laplace”.

Raises

NotImplementedError – When kernel_type is not “Laplace”.

__call__(img)[source]
Parameters

img (Tensor) – torch tensor data to extract the contour, with shape: [batch_size, channels, height, width[, depth]]

Raises

ValueError – When image ndim is not one of [4, 5].

Returns

1. it’s the binary classification result of whether a pixel is edge or not.

2. in order to keep the original shape of mask image, we use padding as default.

3. the edge detection is just approximate because it defects inherent to Laplace kernel, ideally the edge should be thin enough, but now it has a thickness.

Return type

A torch tensor with the same shape as img, note

#### MeanEnsemble¶

class monai.transforms.MeanEnsemble(weights=None)[source]

Execute mean ensemble on the input data. The input data can be a list or tuple of PyTorch Tensor with shape: [B, C[, H, W, D]], Or a single PyTorch Tensor with shape: [E, B, C[, H, W, D]], the E dimension represents the output data from different models. Typically, the input data is model output of segmentation task or classification task. And it also can support to add weights for the input data.

Parameters

weights (Union[Sequence[float], Tensor, ndarray, None]) – can be a list or tuple of numbers for input data with shape: [E, B, C, H, W[, D]]. or a Numpy ndarray or a PyTorch Tensor data. the weights will be added to input data from highest dimension, for example: 1. if the weights only has 1 dimension, it will be added to the E dimension of input data. 2. if the weights has 3 dimensions, it will be added to E, B and C dimensions. it’s a typical practice to add weights for different classes: to ensemble 3 segmentation model outputs, every output has 4 channels(classes), so the input data shape can be: [3, B, 4, H, W, D]. and add different weights for different classes, so the weights shape can be: [3, 1, 4]. for example: weights = [[[1, 2, 3, 4]], [[4, 3, 2, 1]], [[1, 1, 1, 1]]].

__call__(img)[source]

data is an element which often comes from an iteration over an iterable, such as torch.utils.data.Dataset. This method should return an updated version of data. To simplify the input validations, most of the transforms assume that

• data is a Numpy ndarray, PyTorch Tensor or string

• the data shape can be:

1. string data without shape, LoadImage transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChannel expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirst expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

This method can optionally take additional arguments to help execute transformation operation.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Tensor

#### VoteEnsemble¶

class monai.transforms.VoteEnsemble(num_classes=None)[source]

Execute vote ensemble on the input data. The input data can be a list or tuple of PyTorch Tensor with shape: [B[, C, H, W, D]], Or a single PyTorch Tensor with shape: [E, B[, C, H, W, D]], the E dimension represents the output data from different models. Typically, the input data is model output of segmentation task or classification task.

Note

This vote transform expects the input data is discrete values. It can be multiple channels data in One-Hot format or single channel data. It will vote to select the most common data between items. The output data has the same shape as every item of the input data.

Parameters

num_classes (Optional[int]) – if the input is single channel data instead of One-Hot, we can’t get class number from channel, need to explicitly specify the number of classes to vote.

__call__(img)[source]

data is an element which often comes from an iteration over an iterable, such as torch.utils.data.Dataset. This method should return an updated version of data. To simplify the input validations, most of the transforms assume that

• data is a Numpy ndarray, PyTorch Tensor or string

• the data shape can be:

1. string data without shape, LoadImage transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChannel expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirst expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

This method can optionally take additional arguments to help execute transformation operation.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Tensor

### Spatial¶

#### Spacing¶

Resample input image into the specified pixdim.

Parameters
• pixdim (Union[Sequence[float], float]) – output voxel spacing.

• diagonal (bool) –

whether to resample the input to have a diagonal affine matrix. If True, the input data is resampled to the following affine:

np.diag((pixdim_0, pixdim_1, ..., pixdim_n, 1))

This effectively resets the volume to the world coordinate system (RAS+ in nibabel). The original orientation, rotation, shearing are not preserved.

If False, this transform preserves the axes orientation, orthogonal rotation and translation components from the original affine. This option will not flip/swap axes of the original data.

• mode (Union[GridSampleMode, str]) – {"bilinear", "nearest"} Interpolation mode to calculate output values. Defaults to "bilinear". See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample

• align_corners (bool) – Geometrically, we consider the pixels of the input as squares rather than points. See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample

• dtype (Optional[dtype]) – data type for resampling computation. Defaults to np.float64 for best precision. If None, use the data type of input data. To be compatible with other modules, the output data type is always np.float32.

__call__(data_array, affine=None, mode=None, padding_mode=None, align_corners=None, dtype=None)[source]
Parameters
• data_array (ndarray) – in shape (num_channels, H[, W, …]).

• affine (matrix) – (N+1)x(N+1) original affine matrix for spatially ND data_array. Defaults to identity.

• mode (Union[GridSampleMode, str, None]) – {"bilinear", "nearest"} Interpolation mode to calculate output values. Defaults to self.mode. See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample

• align_corners (Optional[bool]) – Geometrically, we consider the pixels of the input as squares rather than points. See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample

• dtype (Optional[dtype]) – data type for resampling computation. Defaults to self.dtype. If None, use the data type of input data. To be compatible with other modules, the output data type is always np.float32.

Raises
• ValueError – When data_array has no spatial dimensions.

• ValueError – When pixdim is nonpositive.

Return type

Tuple[ndarray, ndarray, ndarray]

Returns

data_array (resampled into self.pixdim), original pixdim, current pixdim.

#### Orientation¶

class monai.transforms.Orientation(axcodes=None, as_closest_canonical=False, labels=(('L', 'R'), ('P', 'A'), ('I', 'S')))[source]

Change the input image’s orientation into the specified based on axcodes.

Parameters
• axcodes (Optional[str]) – N elements sequence for spatial ND input’s orientation. e.g. axcodes=’RAS’ represents 3D orientation: (Left, Right), (Posterior, Anterior), (Inferior, Superior). default orientation labels options are: ‘L’ and ‘R’ for the first dimension, ‘P’ and ‘A’ for the second, ‘I’ and ‘S’ for the third.

• as_closest_canonical (bool) – if True, load the image as closest to canonical axis format.

• labels (Optional[Sequence[Tuple[str, str]]]) – optional, None or sequence of (2,) sequences (2,) sequences are labels for (beginning, end) of output axis. Defaults to (('L', 'R'), ('P', 'A'), ('I', 'S')).

Raises

ValueError – When axcodes=None and as_closest_canonical=True. Incompatible values.

__call__(data_array, affine=None)[source]

original orientation of data_array is defined by affine.

Parameters
• data_array (ndarray) – in shape (num_channels, H[, W, …]).

• affine (matrix) – (N+1)x(N+1) original affine matrix for spatially ND data_array. Defaults to identity.

Raises
• ValueError – When data_array has no spatial dimensions.

• ValueError – When axcodes spatiality differs from data_array.

Return type

Tuple[ndarray, ndarray, ndarray]

Returns

data_array (reoriented in self.axcodes), original axcodes, current axcodes.

#### RandRotate¶

class monai.transforms.RandRotate(range_x=0.0, range_y=0.0, range_z=0.0, prob=0.1, keep_size=True, mode=<GridSampleMode.BILINEAR: 'bilinear'>, padding_mode=<GridSamplePadMode.BORDER: 'border'>, align_corners=False, dtype=<class 'numpy.float64'>)[source]

Randomly rotate the input arrays.

Parameters
• range_x (Union[Tuple[float, float], float]) – Range of rotation angle in radians in the plane defined by the first and second axes. If single number, angle is uniformly sampled from (-range_x, range_x).

• range_y (Union[Tuple[float, float], float]) – Range of rotation angle in radians in the plane defined by the first and third axes. If single number, angle is uniformly sampled from (-range_y, range_y).

• range_z (Union[Tuple[float, float], float]) – Range of rotation angle in radians in the plane defined by the second and third axes. If single number, angle is uniformly sampled from (-range_z, range_z).

• prob (float) – Probability of rotation.

• keep_size (bool) – If it is False, the output shape is adapted so that the input array is contained completely in the output. If it is True, the output shape is the same as the input. Default is True.

• mode (Union[GridSampleMode, str]) – {"bilinear", "nearest"} Interpolation mode to calculate output values. Defaults to "bilinear". See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample

• dtype (Optional[dtype]) – data type for resampling computation. Defaults to np.float64 for best precision. If None, use the data type of input data. To be compatible with other modules, the output data type is always np.float32.

Parameters
Return type

ndarray

randomize(data=None)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

#### RandFlip¶

class monai.transforms.RandFlip(prob=0.1, spatial_axis=None)[source]

Randomly flips the image along axes. Preserves shape. See numpy.flip for additional details. https://docs.scipy.org/doc/numpy/reference/generated/numpy.flip.html

Parameters
• prob (float) – Probability of flipping.

• spatial_axis (Union[Sequence[int], int, None]) – Spatial axes along which to flip over. Default is None.

__call__(img)[source]
Parameters

img (ndarray) – channel first array, must have shape: (num_channels, H[, W, …, ]),

Return type

ndarray

randomize(data=None)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

#### RandZoom¶

Randomly zooms input arrays with given probability within given zoom range.

Parameters
• prob (float) – Probability of zooming.

• min_zoom (Union[Sequence[float], float]) – Min zoom factor. Can be float or sequence same size as image. If a float, select a random factor from [min_zoom, max_zoom] then apply to all spatial dims to keep the original spatial shape ratio. If a sequence, min_zoom should contain one value for each spatial axis. If 2 values provided for 3D data, use the first value for both H & W dims to keep the same zoom ratio.

• max_zoom (Union[Sequence[float], float]) – Max zoom factor. Can be float or sequence same size as image. If a float, select a random factor from [min_zoom, max_zoom] then apply to all spatial dims to keep the original spatial shape ratio. If a sequence, max_zoom should contain one value for each spatial axis. If 2 values provided for 3D data, use the first value for both H & W dims to keep the same zoom ratio.

• mode (Union[InterpolateMode, str]) – {"nearest", "linear", "bilinear", "bicubic", "trilinear", "area"} The interpolation mode. Defaults to "area". See also: https://pytorch.org/docs/stable/nn.functional.html#interpolate

• padding_mode (Union[NumpyPadMode, str]) – {"constant", "edge”, "linear_ramp”, "maximum”, "mean”, “median”, "minimum”, “reflect”, "symmetric”, "wrap”, "empty”, "<function>”} The mode to pad data after zooming. See also: https://numpy.org/doc/stable/reference/generated/numpy.pad.html

• align_corners (Optional[bool]) – This only has an effect when mode is ‘linear’, ‘bilinear’, ‘bicubic’ or ‘trilinear’. Default: None. See also: https://pytorch.org/docs/stable/nn.functional.html#interpolate

• keep_size (bool) – Should keep original size (pad if needed), default is True.

Parameters
Return type

ndarray

randomize(data=None)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

#### Affine¶

class monai.transforms.Affine(rotate_params=None, shear_params=None, translate_params=None, scale_params=None, spatial_size=None, mode=<GridSampleMode.BILINEAR: 'bilinear'>, padding_mode=<GridSamplePadMode.REFLECTION: 'reflection'>, as_tensor_output=False, device=None)[source]

Transform img given the affine parameters.

The affine transformations are applied in rotate, shear, translate, scale order.

Parameters
• rotate_params (Union[Sequence[float], float, None]) – a rotation angle in radians, a scalar for 2D image, a tuple of 3 floats for 3D. Defaults to no rotation.

• shear_params (Union[Sequence[float], float, None]) – a tuple of 2 floats for 2D, a tuple of 6 floats for 3D. Defaults to no shearing.

• translate_params (Union[Sequence[float], float, None]) – a tuple of 2 floats for 2D, a tuple of 3 floats for 3D. Translation is in pixel/voxel relative to the center of the input image. Defaults to no translation.

• scale_params (Union[Sequence[float], float, None]) – a tuple of 2 floats for 2D, a tuple of 3 floats for 3D. Defaults to no scaling.

• spatial_size (Union[Sequence[int], int, None]) – output image spatial size. if spatial_size and self.spatial_size are not defined, or smaller than 1, the transform will use the spatial size of img. if the components of the spatial_size are non-positive values, the transform will use the corresponding components of img size. For example, spatial_size=(32, -1) will be adapted to (32, 64) if the second spatial dimension size of img is 64.

• mode (Union[GridSampleMode, str]) – {"bilinear", "nearest"} Interpolation mode to calculate output values. Defaults to "bilinear". See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample

• as_tensor_output (bool) – the computation is implemented using pytorch tensors, this option specifies whether to convert it back to numpy arrays.

• device (Optional[device]) – device on which the tensor will be allocated.

Parameters
• img (Union[ndarray, Tensor]) – shape must be (num_channels, H, W[, D]),

• spatial_size (Union[Sequence[int], int, None]) – output image spatial size. if spatial_size and self.spatial_size are not defined, or smaller than 1, the transform will use the spatial size of img. if img has two spatial dimensions, spatial_size should have 2 elements [h, w]. if img has three spatial dimensions, spatial_size should have 3 elements [h, w, d].

• mode (Union[GridSampleMode, str, None]) – {"bilinear", "nearest"} Interpolation mode to calculate output values. Defaults to self.mode. See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample

Return type

Union[ndarray, Tensor]

#### Resample¶

computes output image using values from img, locations from grid using pytorch. supports spatially 2D or 3D (num_channels, H, W[, D]).

Parameters
Parameters
Return type

Union[ndarray, Tensor]

#### RandAffine¶

class monai.transforms.RandAffine(prob=0.1, rotate_range=None, shear_range=None, translate_range=None, scale_range=None, spatial_size=None, mode=<GridSampleMode.BILINEAR: 'bilinear'>, padding_mode=<GridSamplePadMode.REFLECTION: 'reflection'>, as_tensor_output=True, device=None)[source]

Random affine transform.

Parameters
• prob (float) – probability of returning a randomized affine grid. defaults to 0.1, with 10% chance returns a randomized grid.

• rotate_range (Union[Sequence[float], float, None]) – angle range in radians. rotate_range[0] with be used to generate the 1st rotation parameter from uniform[-rotate_range[0], rotate_range[0]). Similarly, rotate_range[1] and rotate_range[2] are used in 3D affine for the range of 2nd and 3rd axes.

• shear_range (Union[Sequence[float], float, None]) – shear_range[0] with be used to generate the 1st shearing parameter from uniform[-shear_range[0], shear_range[0]). Similarly, shear_range[1] to shear_range[N] controls the range of the uniform distribution used to generate the 2nd to N-th parameter.

• translate_range (Union[Sequence[float], float, None]) – translate_range[0] with be used to generate the 1st shift parameter from uniform[-translate_range[0], translate_range[0]). Similarly, translate_range[1] to translate_range[N] controls the range of the uniform distribution used to generate the 2nd to N-th parameter.

• scale_range (Union[Sequence[float], float, None]) – scaling_range[0] with be used to generate the 1st scaling factor from uniform[-scale_range[0], scale_range[0]) + 1.0. Similarly, scale_range[1] to scale_range[N] controls the range of the uniform distribution used to generate the 2nd to N-th parameter.

• spatial_size (Union[Sequence[float], float, None]) – output image spatial size. if spatial_size and self.spatial_size are not defined, or smaller than 1, the transform will use the spatial size of img. if the components of the spatial_size are non-positive values, the transform will use the corresponding components of img size. For example, spatial_size=(32, -1) will be adapted to (32, 64) if the second spatial dimension size of img is 64.

• mode (Union[GridSampleMode, str]) – {"bilinear", "nearest"} Interpolation mode to calculate output values. Defaults to "bilinear". See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample

• as_tensor_output (bool) – the computation is implemented using pytorch tensors, this option specifies whether to convert it back to numpy arrays.

• device (Optional[device]) – device on which the tensor will be allocated.

• RandAffineGrid for the random affine parameters configurations.

• Affine for the affine transformation parameters configurations.

Parameters
• img (Union[ndarray, Tensor]) – shape must be (num_channels, H, W[, D]),

• spatial_size (Union[Sequence[int], int, None]) – output image spatial size. if spatial_size and self.spatial_size are not defined, or smaller than 1, the transform will use the spatial size of img. if img has two spatial dimensions, spatial_size should have 2 elements [h, w]. if img has three spatial dimensions, spatial_size should have 3 elements [h, w, d].

• mode (Union[GridSampleMode, str, None]) – {"bilinear", "nearest"} Interpolation mode to calculate output values. Defaults to self.mode. See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample

Return type

Union[ndarray, Tensor]

randomize(data=None)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

set_random_state(seed=None, state=None)[source]

Set the random state locally, to control the randomness, the derived classes should use self.R instead of np.random to introduce random factors.

Parameters
• seed (Optional[int]) – set the random state with an integer seed.

• state (Optional[RandomState]) – set the random state with a np.random.RandomState object.

Raises

TypeError – When state is not an Optional[np.random.RandomState].

Return type

RandAffine

Returns

a Randomizable instance.

#### RandDeformGrid¶

class monai.transforms.RandDeformGrid(spacing, magnitude_range, as_tensor_output=True, device=None)[source]

Generate random deformation grid.

Parameters
• spacing (Union[Sequence[float], float]) – spacing of the grid in 2D or 3D. e.g., spacing=(1, 1) indicates pixel-wise deformation in 2D, spacing=(1, 1, 1) indicates voxel-wise deformation in 3D, spacing=(2, 2) indicates deformation field defined on every other pixel in 2D.

• magnitude_range (Tuple[float, float]) – the random offsets will be generated from uniform[magnitude[0], magnitude[1]).

• as_tensor_output (bool) – whether to output tensor instead of numpy array. defaults to True.

• device (Optional[device]) – device to store the output grid data.

__call__(spatial_size)[source]
Parameters

spatial_size (Sequence[int]) – spatial size of the grid.

Return type

Union[ndarray, Tensor]

randomize(grid_size)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

#### AffineGrid¶

class monai.transforms.AffineGrid(rotate_params=None, shear_params=None, translate_params=None, scale_params=None, as_tensor_output=True, device=None)[source]

Affine transforms on the coordinates.

Parameters
• rotate_params (Union[Sequence[float], float, None]) – angle range in radians. rotate_params[0] with be used to generate the 1st rotation parameter from uniform[-rotate_params[0], rotate_params[0]). Similarly, rotate_params[1] and rotate_params[2] are used in 3D affine for the range of 2nd and 3rd axes.

• shear_params (Union[Sequence[float], float, None]) – shear_params[0] with be used to generate the 1st shearing parameter from uniform[-shear_params[0], shear_params[0]). Similarly, shear_params[1] to shear_params[N] controls the range of the uniform distribution used to generate the 2nd to N-th parameter.

• translate_params (Union[Sequence[float], float, None]) – translate_params[0] with be used to generate the 1st shift parameter from uniform[-translate_params[0], translate_params[0]). Similarly, translate_params[1] to translate_params[N] controls the range of the uniform distribution used to generate the 2nd to N-th parameter.

• scale_params (Union[Sequence[float], float, None]) – scale_params[0] with be used to generate the 1st scaling factor from uniform[-scale_params[0], scale_params[0]) + 1.0. Similarly, scale_params[1] to scale_params[N] controls the range of the uniform distribution used to generate the 2nd to N-th parameter.

• as_tensor_output (bool) – whether to output tensor instead of numpy array. defaults to True.

• device (Optional[device]) – device to store the output grid data.

__call__(spatial_size=None, grid=None)[source]
Parameters
• spatial_size (Optional[Sequence[int]]) – output grid size.

• grid (Union[ndarray, Tensor, None]) – grid to be transformed. Shape must be (3, H, W) for 2D or (4, H, W, D) for 3D.

Raises

ValueError – When grid=None and spatial_size=None. Incompatible values.

Return type

Union[ndarray, Tensor]

#### RandAffineGrid¶

class monai.transforms.RandAffineGrid(rotate_range=None, shear_range=None, translate_range=None, scale_range=None, as_tensor_output=True, device=None)[source]

Generate randomised affine grid.

Parameters
• rotate_range (Union[Sequence[float], float, None]) – angle range in radians. rotate_range[0] with be used to generate the 1st rotation parameter from uniform[-rotate_range[0], rotate_range[0]). Similarly, rotate_range[1] and rotate_range[2] are used in 3D affine for the range of 2nd and 3rd axes.

• shear_range (Union[Sequence[float], float, None]) – shear_range[0] with be used to generate the 1st shearing parameter from uniform[-shear_range[0], shear_range[0]). Similarly, shear_range[1] to shear_range[N] controls the range of the uniform distribution used to generate the 2nd to N-th parameter.

• translate_range (Union[Sequence[float], float, None]) – translate_range[0] with be used to generate the 1st shift parameter from uniform[-translate_range[0], translate_range[0]). Similarly, translate_range[1] to translate_range[N] controls the range of the uniform distribution used to generate the 2nd to N-th parameter.

• scale_range (Union[Sequence[float], float, None]) – scaling_range[0] with be used to generate the 1st scaling factor from uniform[-scale_range[0], scale_range[0]) + 1.0. Similarly, scale_range[1] to scale_range[N] controls the range of the uniform distribution used to generate the 2nd to N-th parameter.

• as_tensor_output (bool) – whether to output tensor instead of numpy array. defaults to True.

• device (Optional[device]) – device to store the output grid data.

__call__(spatial_size=None, grid=None)[source]
Parameters
• spatial_size (Optional[Sequence[int]]) – output grid size.

• grid (Union[ndarray, Tensor, None]) – grid to be transformed. Shape must be (3, H, W) for 2D or (4, H, W, D) for 3D.

Return type

Union[ndarray, Tensor]

Returns

a 2D (3xHxW) or 3D (4xHxWxD) grid.

randomize(data=None)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

#### Rand2DElastic¶

class monai.transforms.Rand2DElastic(spacing, magnitude_range, prob=0.1, rotate_range=None, shear_range=None, translate_range=None, scale_range=None, spatial_size=None, mode=<GridSampleMode.BILINEAR: 'bilinear'>, padding_mode=<GridSamplePadMode.REFLECTION: 'reflection'>, as_tensor_output=False, device=None)[source]

Random elastic deformation and affine in 2D

Parameters
• spacing (Union[Tuple[float, float], float]) – distance in between the control points.

• magnitude_range (Tuple[float, float]) – the random offsets will be generated from uniform[magnitude[0], magnitude[1]).

• prob (float) – probability of returning a randomized elastic transform. defaults to 0.1, with 10% chance returns a randomized elastic transform, otherwise returns a spatial_size centered area extracted from the input image.

• rotate_range (Union[Sequence[float], float, None]) – angle range in radians. rotate_range[0] with be used to generate the 1st rotation parameter from uniform[-rotate_range[0], rotate_range[0]).

• shear_range (Union[Sequence[float], float, None]) – shear_range[0] with be used to generate the 1st shearing parameter from uniform[-shear_range[0], shear_range[0]). Similarly, shear_range[1] controls the range of the uniform distribution used to generate the 2nd parameter.

• translate_range (Union[Sequence[float], float, None]) – translate_range[0] with be used to generate the 1st shift parameter from uniform[-translate_range[0], translate_range[0]). Similarly, translate_range[1] controls the range of the uniform distribution used to generate the 2nd parameter.

• scale_range (Union[Sequence[float], float, None]) – scaling_range[0] with be used to generate the 1st scaling factor from uniform[-scale_range[0], scale_range[0]) + 1.0. Similarly, scale_range[1] controls the range of the uniform distribution used to generate the 2nd parameter.

• spatial_size (Union[Sequence[int], int, None]) – specifying output image spatial size [h, w]. if spatial_size and self.spatial_size are not defined, or smaller than 1, the transform will use the spatial size of img. if the components of the spatial_size are non-positive values, the transform will use the corresponding components of img size. For example, spatial_size=(32, -1) will be adapted to (32, 64) if the second spatial dimension size of img is 64.

• mode (Union[GridSampleMode, str]) – {"bilinear", "nearest"} Interpolation mode to calculate output values. Defaults to "bilinear". See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample

• as_tensor_output (bool) – the computation is implemented using pytorch tensors, this option specifies whether to convert it back to numpy arrays.

• device (Optional[device]) – device on which the tensor will be allocated.

• RandAffineGrid for the random affine parameters configurations.

• Affine for the affine transformation parameters configurations.

Parameters
• img (Union[ndarray, Tensor]) – shape must be (num_channels, H, W),

• spatial_size (Union[Tuple[int, int], int, None]) – specifying output image spatial size [h, w]. if spatial_size and self.spatial_size are not defined, or smaller than 1, the transform will use the spatial size of img.

• mode (Union[GridSampleMode, str, None]) – {"bilinear", "nearest"} Interpolation mode to calculate output values. Defaults to self.mode. See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample

Return type

Union[ndarray, Tensor]

randomize(spatial_size)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

set_random_state(seed=None, state=None)[source]

Set the random state locally, to control the randomness, the derived classes should use self.R instead of np.random to introduce random factors.

Parameters
• seed (Optional[int]) – set the random state with an integer seed.

• state (Optional[RandomState]) – set the random state with a np.random.RandomState object.

Raises

TypeError – When state is not an Optional[np.random.RandomState].

Return type

Rand2DElastic

Returns

a Randomizable instance.

#### Rand3DElastic¶

class monai.transforms.Rand3DElastic(sigma_range, magnitude_range, prob=0.1, rotate_range=None, shear_range=None, translate_range=None, scale_range=None, spatial_size=None, mode=<GridSampleMode.BILINEAR: 'bilinear'>, padding_mode=<GridSamplePadMode.REFLECTION: 'reflection'>, as_tensor_output=False, device=None)[source]

Random elastic deformation and affine in 3D

Parameters
• sigma_range (Tuple[float, float]) – a Gaussian kernel with standard deviation sampled from uniform[sigma_range[0], sigma_range[1]) will be used to smooth the random offset grid.

• magnitude_range (Tuple[float, float]) – the random offsets on the grid will be generated from uniform[magnitude[0], magnitude[1]).

• prob (float) – probability of returning a randomized elastic transform. defaults to 0.1, with 10% chance returns a randomized elastic transform, otherwise returns a spatial_size centered area extracted from the input image.

• rotate_range (Union[Sequence[float], float, None]) – angle range in radians. rotate_range[0] with be used to generate the 1st rotation parameter from uniform[-rotate_range[0], rotate_range[0]). Similarly, rotate_range[1] and rotate_range[2] are used in 3D affine for the range of 2nd and 3rd axes.

• shear_range (Union[Sequence[float], float, None]) – shear_range[0] with be used to generate the 1st shearing parameter from uniform[-shear_range[0], shear_range[0]). Similarly, shear_range[1] and shear_range[2] controls the range of the uniform distribution used to generate the 2nd and 3rd parameters.

• translate_range (Union[Sequence[float], float, None]) – translate_range[0] with be used to generate the 1st shift parameter from uniform[-translate_range[0], translate_range[0]). Similarly, translate_range[1] and translate_range[2] controls the range of the uniform distribution used to generate the 2nd and 3rd parameters.

• scale_range (Union[Sequence[float], float, None]) – scaling_range[0] with be used to generate the 1st scaling factor from uniform[-scale_range[0], scale_range[0]) + 1.0. Similarly, scale_range[1] and scale_range[2] controls the range of the uniform distribution used to generate the 2nd and 3rd parameters.

• spatial_size (Union[Sequence[int], int, None]) – specifying output image spatial size [h, w, d]. if spatial_size and self.spatial_size are not defined, or smaller than 1, the transform will use the spatial size of img. if the components of the spatial_size are non-positive values, the transform will use the corresponding components of img size. For example, spatial_size=(32, 32, -1) will be adapted to (32, 32, 64) if the third spatial dimension size of img is 64.

• mode (Union[GridSampleMode, str]) – {"bilinear", "nearest"} Interpolation mode to calculate output values. Defaults to "bilinear". See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample

• as_tensor_output (bool) – the computation is implemented using pytorch tensors, this option specifies whether to convert it back to numpy arrays.

• device (Optional[device]) – device on which the tensor will be allocated.

• RandAffineGrid for the random affine parameters configurations.

• Affine for the affine transformation parameters configurations.

Parameters
• img (Union[ndarray, Tensor]) – shape must be (num_channels, H, W, D),

• spatial_size (Union[Tuple[int, int, int], int, None]) – specifying spatial 3D output image spatial size [h, w, d]. if spatial_size and self.spatial_size are not defined, or smaller than 1, the transform will use the spatial size of img.

• mode (Union[GridSampleMode, str, None]) – {"bilinear", "nearest"} Interpolation mode to calculate output values. Defaults to self.mode. See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample

Return type

Union[ndarray, Tensor]

randomize(grid_size)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

set_random_state(seed=None, state=None)[source]

Set the random state locally, to control the randomness, the derived classes should use self.R instead of np.random to introduce random factors.

Parameters
• seed (Optional[int]) – set the random state with an integer seed.

• state (Optional[RandomState]) – set the random state with a np.random.RandomState object.

Raises

TypeError – When state is not an Optional[np.random.RandomState].

Return type

Rand3DElastic

Returns

a Randomizable instance.

#### Rotate90¶

class monai.transforms.Rotate90(k=1, spatial_axes=(0, 1))[source]

Rotate an array by 90 degrees in the plane specified by axes.

Parameters
• k (int) – number of times to rotate by 90 degrees.

• spatial_axes (Tuple[int, int]) – 2 int numbers, defines the plane to rotate with 2 spatial axes. Default: (0, 1), this is the first two axis in spatial dimensions.

__call__(img)[source]
Parameters

img (ndarray) – channel first array, must have shape: (num_channels, H[, W, …, ]),

Return type

ndarray

#### RandRotate90¶

class monai.transforms.RandRotate90(prob=0.1, max_k=3, spatial_axes=(0, 1))[source]

With probability prob, input arrays are rotated by 90 degrees in the plane specified by spatial_axes.

Parameters
• prob (float) – probability of rotating. (Default 0.1, with 10% probability it returns a rotated array)

• max_k (int) – number of rotations will be sampled from np.random.randint(max_k) + 1, (Default 3).

• spatial_axes (Tuple[int, int]) – 2 int numbers, defines the plane to rotate with 2 spatial axes. Default: (0, 1), this is the first two axis in spatial dimensions.

__call__(img)[source]
Parameters

img (ndarray) – channel first array, must have shape: (num_channels, H[, W, …, ]),

Return type

ndarray

randomize(data=None)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

#### Flip¶

class monai.transforms.Flip(spatial_axis)[source]

Reverses the order of elements along the given spatial axis. Preserves shape. Uses np.flip in practice. See numpy.flip for additional details. https://docs.scipy.org/doc/numpy/reference/generated/numpy.flip.html

Parameters

spatial_axis (Union[Sequence[int], int, None]) – spatial axes along which to flip over. Default is None.

__call__(img)[source]
Parameters

img (ndarray) – channel first array, must have shape: (num_channels, H[, W, …, ]),

Return type

ndarray

#### Resize¶

class monai.transforms.Resize(spatial_size, mode=<InterpolateMode.AREA: 'area'>, align_corners=None)[source]

Resize the input image to given spatial size. Implemented using torch.nn.functional.interpolate.

Parameters
• spatial_size (Union[Sequence[int], int]) – expected shape of spatial dimensions after resize operation. if the components of the spatial_size are non-positive values, the transform will use the corresponding components of img size. For example, spatial_size=(32, -1) will be adapted to (32, 64) if the second spatial dimension size of img is 64.

• mode (Union[InterpolateMode, str]) – {"nearest", "linear", "bilinear", "bicubic", "trilinear", "area"} The interpolation mode. Defaults to "area". See also: https://pytorch.org/docs/stable/nn.functional.html#interpolate

• align_corners (Optional[bool]) – This only has an effect when mode is ‘linear’, ‘bilinear’, ‘bicubic’ or ‘trilinear’. Default: None. See also: https://pytorch.org/docs/stable/nn.functional.html#interpolate

__call__(img, mode=None, align_corners=None)[source]
Parameters
Raises

ValueError – When self.spatial_size length is less than img spatial dimensions.

Return type

ndarray

#### Rotate¶

Rotates an input image by given angle using monai.networks.layers.AffineTransform.

Parameters
• angle (Union[Sequence[float], float]) – Rotation angle(s) in radians. should a float for 2D, three floats for 3D.

• keep_size (bool) – If it is True, the output shape is kept the same as the input. If it is False, the output shape is adapted so that the input array is contained completely in the output. Default is True.

• mode (Union[GridSampleMode, str]) – {"bilinear", "nearest"} Interpolation mode to calculate output values. Defaults to "bilinear". See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample

• dtype (Optional[dtype]) – data type for resampling computation. Defaults to np.float64 for best precision. If None, use the data type of input data. To be compatible with other modules, the output data type is always np.float32.

Parameters
Raises

ValueError – When img spatially is not one of [2D, 3D].

Return type

ndarray

#### Zoom¶

Zooms an ND image using torch.nn.functional.interpolate. For details, please see https://pytorch.org/docs/stable/nn.functional.html#interpolate.

Different from monai.transforms.resize, this transform takes scaling factors as input, and provides an option of preserving the input spatial size.

Parameters
• zoom (Union[Sequence[float], float]) – The zoom factor along the spatial axes. If a float, zoom is the same for each spatial axis. If a sequence, zoom should contain one value for each spatial axis.

• mode (Union[InterpolateMode, str]) – {"nearest", "linear", "bilinear", "bicubic", "trilinear", "area"} The interpolation mode. Defaults to "area". See also: https://pytorch.org/docs/stable/nn.functional.html#interpolate

• padding_mode (Union[NumpyPadMode, str]) – {"constant", "edge”, "linear_ramp”, "maximum”, "mean”, “median”, "minimum”, “reflect”, "symmetric”, "wrap”, "empty”, "<function>”} The mode to pad data after zooming. See also: https://numpy.org/doc/stable/reference/generated/numpy.pad.html

• align_corners (Optional[bool]) – This only has an effect when mode is ‘linear’, ‘bilinear’, ‘bicubic’ or ‘trilinear’. Default: None. See also: https://pytorch.org/docs/stable/nn.functional.html#interpolate

• keep_size (bool) – Should keep original size (padding/slicing if needed), default is True.

Parameters
Return type

ndarray

### Utility¶

#### Identity¶

class monai.transforms.Identity[source]

Convert the input to an np.ndarray, if input data is np.ndarray or subclasses, return unchanged data. As the output value is same as input, it can be used as a testing tool to verify the transform chain, Compose or transform adaptor, etc.

__call__(img)[source]

Apply the transform to img.

Return type

ndarray

#### AsChannelFirst¶

class monai.transforms.AsChannelFirst(channel_dim=- 1)[source]

Change the channel dimension of the image to the first dimension.

Most of the image transformations in monai.transforms assume the input image is in the channel-first format, which has the shape (num_channels, spatial_dim_1[, spatial_dim_2, …]).

This transform could be used to convert, for example, a channel-last image array in shape (spatial_dim_1[, spatial_dim_2, …], num_channels) into the channel-first format, so that the multidimensional image array can be correctly interpreted by the other transforms.

Parameters

channel_dim (int) – which dimension of input image is the channel, default is the last dimension.

__call__(img)[source]

Apply the transform to img.

Return type

ndarray

#### AsChannelLast¶

class monai.transforms.AsChannelLast(channel_dim=0)[source]

Change the channel dimension of the image to the last dimension.

Some of other 3rd party transforms assume the input image is in the channel-last format with shape (spatial_dim_1[, spatial_dim_2, …], num_channels).

This transform could be used to convert, for example, a channel-first image array in shape (num_channels, spatial_dim_1[, spatial_dim_2, …]) into the channel-last format, so that MONAI transforms can construct a chain with other 3rd party transforms together.

Parameters

channel_dim (int) – which dimension of input image is the channel, default is the first dimension.

__call__(img)[source]

Apply the transform to img.

Return type

ndarray

Adds a 1-length channel dimension to the input image.

Most of the image transformations in monai.transforms assumes the input image is in the channel-first format, which has the shape (num_channels, spatial_dim_1[, spatial_dim_2, …]).

This transform could be used, for example, to convert a (spatial_dim_1[, spatial_dim_2, …]) spatial image into the channel-first format so that the multidimensional image array can be correctly interpreted by the other transforms.

__call__(img)[source]

Apply the transform to img.

Return type

~NdarrayTensor

#### RepeatChannel¶

class monai.transforms.RepeatChannel(repeats)[source]

Repeat channel data to construct expected input shape for models. The repeats count includes the origin data, for example: RepeatChannel(repeats=2)([[1, 2], [3, 4]]) generates: [[1, 2], [1, 2], [3, 4], [3, 4]]

Parameters

repeats (int) – the number of repetitions for each element.

__call__(img)[source]

Apply the transform to img, assuming img is a “channel-first” array.

Return type

ndarray

#### SplitChannel¶

class monai.transforms.SplitChannel(channel_dim=None)[source]

Split Numpy array or PyTorch Tensor data according to the channel dim. It can help applying different following transforms to different channels. Channel number must be greater than 1.

Parameters

channel_dim (Optional[int]) – which dimension of input image is the channel, default to None to automatically select: if data is numpy array, channel_dim is 0 as numpy array is used in the pre transforms, if PyTorch Tensor, channel_dim is 1 as in most of the cases Tensor is uses in the post transforms.

__call__(img)[source]

data is an element which often comes from an iteration over an iterable, such as torch.utils.data.Dataset. This method should return an updated version of data. To simplify the input validations, most of the transforms assume that

• data is a Numpy ndarray, PyTorch Tensor or string

• the data shape can be:

1. string data without shape, LoadImage transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChannel expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirst expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

This method can optionally take additional arguments to help execute transformation operation.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

List[Union[ndarray, Tensor]]

#### CastToType¶

class monai.transforms.CastToType(dtype=<class 'numpy.float32'>)[source]

Cast the Numpy data to specified numpy data type, or cast the PyTorch Tensor to specified PyTorch data type.

Parameters

dtype (Union[dtype, dtype]) – convert image to this data type, default is np.float32.

__call__(img, dtype=None)[source]

Apply the transform to img, assuming img is a numpy array or PyTorch Tensor.

Parameters

dtype (Union[dtype, dtype, None]) – convert image to this data type, default is self.dtype.

Raises

TypeError – When img type is not in Union[numpy.ndarray, torch.Tensor].

Return type

Union[ndarray, Tensor]

#### ToTensor¶

class monai.transforms.ToTensor[source]

Converts the input image to a tensor without applying any other transformations.

__call__(img)[source]

Apply the transform to img and make it contiguous.

Return type

Tensor

#### ToNumpy¶

class monai.transforms.ToNumpy[source]

Converts the input data to numpy array, can support list or tuple of numbers and PyTorch Tensor.

__call__(img)[source]

Apply the transform to img and make it contiguous.

Return type

ndarray

#### Transpose¶

class monai.transforms.Transpose(indices)[source]

Transposes the input image based on the given indices dimension ordering.

__call__(img)[source]

Apply the transform to img.

Return type

ndarray

#### SqueezeDim¶

class monai.transforms.SqueezeDim(dim=0)[source]

Squeeze a unitary dimension.

Parameters

dim (Optional[int]) – dimension to be squeezed. Default = 0 “None” works when the input is numpy array.

Raises

TypeError – When dim is not an Optional[int].

__call__(img)[source]
Parameters

img (~NdarrayTensor) – numpy arrays with required dimension dim removed

Return type

~NdarrayTensor

#### DataStats¶

class monai.transforms.DataStats(prefix='Data', data_shape=True, value_range=True, data_value=False, additional_info=None, logger_handler=None)[source]

Utility transform to show the statistics of data for debug or analysis. It can be inserted into any place of a transform chain and check results of previous transforms. It support both numpy.ndarray and torch.tensor as input data, so it can be used in pre-processing and post-processing.

Parameters
• prefix (str) – will be printed in format: “{prefix} statistics”.

• data_shape (bool) – whether to show the shape of input data.

• value_range (bool) – whether to show the value range of input data.

• data_value (bool) – whether to show the raw value of input data. a typical example is to print some properties of Nifti image: affine, pixdim, etc.

• additional_info (Optional[Callable]) – user can define callable function to extract additional info from input data.

• logger_handler (Optional[Handler]) – add additional handler to output data: save to file, etc. add existing python logging handlers: https://docs.python.org/3/library/logging.handlers.html

Raises

TypeError – When additional_info is not an Optional[Callable].

__call__(img, prefix=None, data_shape=None, value_range=None, data_value=None, additional_info=None)[source]

Apply the transform to img, optionally take arguments similar to the class constructor.

Return type

~NdarrayTensor

#### SimulateDelay¶

class monai.transforms.SimulateDelay(delay_time=0.0)[source]

This is a pass through transform to be used for testing purposes. It allows adding fake behaviors that are useful for testing purposes to simulate how large datasets behave without needing to test on large data sets.

For example, simulating slow NFS data transfers, or slow network transfers in testing by adding explicit timing delays. Testing of small test data can lead to incomplete understanding of real world issues, and may lead to sub-optimal design choices.

Parameters

delay_time (float) – The minimum amount of time, in fractions of seconds, to accomplish this delay task.

__call__(img, delay_time=None)[source]
Parameters
• img (~NdarrayTensor) – data remain unchanged throughout this transform.

• delay_time (Optional[float]) – The minimum amount of time, in fractions of seconds, to accomplish this delay task.

Return type

~NdarrayTensor

#### Lambda¶

class monai.transforms.Lambda(func=None)[source]

Apply a user-defined lambda as a transform.

For example:

image = np.ones((10, 2, 2))
lambd = Lambda(func=lambda x: x[:4, :, :])
print(lambd(image).shape)
(4, 2, 2)
Parameters

func (Optional[Callable]) – Lambda/function to be applied.

Raises

TypeError – When func is not an Optional[Callable].

__call__(img, func=None)[source]

Apply self.func to img.

Parameters

func (Optional[Callable]) – Lambda/function to be applied. Defaults to self.func.

Raises
• TypeError – When func is not an Optional[Callable].

• ValueError – When func=None and self.func=None. Incompatible values.

Convert labels to mask for other tasks. A typical usage is to convert segmentation labels to mask data to pre-process images and then feed the images into classification network. It can support single channel labels or One-Hot labels with specified select_labels. For example, users can select label value = [2, 3] to construct mask data, or select the second and the third channels of labels to construct mask data. The output mask data can be a multiple channels binary data or a single channel binary data that merges all the channels.

Parameters
• select_labels (Union[Sequence[int], int]) – labels to generate mask from. for 1 channel label, the select_labels is the expected label values, like: [1, 2, 3]. for One-Hot format label, the select_labels is the expected channel indices.

• merge_channels (bool) – whether to use np.any() to merge the result on channel dim. if yes, will return a single channel mask with binary data.

__call__(img, select_labels=None, merge_channels=False)[source]
Parameters
• select_labels (Union[Sequence[int], int, None]) – labels to generate mask from. for 1 channel label, the select_labels is the expected label values, like: [1, 2, 3]. for One-Hot format label, the select_labels is the expected channel indices.

• merge_channels (bool) – whether to use np.any() to merge the result on channel dim. if yes, will return a single channel mask with binary data.

Return type

ndarray

#### FgBgToIndices¶

class monai.transforms.FgBgToIndices(image_threshold=0.0, output_shape=None)[source]

Compute foreground and background of the input label data, return the indices. If no output_shape specified, output data will be 1 dim indices after flattening. This transform can help pre-compute foreground and background regions for other transforms. A typical usage is to randomly select foreground and background to crop. The main logic is based on monai.transforms.utils.map_binary_to_indices.

Parameters
• image_threshold (float) – if enabled image at runtime, use image > image_threshold to determine the valid image content area and select background only in this area.

• output_shape (Optional[Sequence[int]]) – expected shape of output indices. if not None, unravel indices to specified shape.

__call__(label, image=None, output_shape=None)[source]
Parameters
• label (ndarray) – input data to compute foreground and background indices.

• image (Optional[ndarray]) – if image is not None, use label = 0 & image > image_threshold to define background. so the output items will not map to all the voxels in the label.

• output_shape (Optional[Sequence[int]]) – expected shape of output indices. if None, use self.output_shape instead.

Return type

Tuple[ndarray, ndarray]

#### ConvertToMultiChannelBasedOnBratsClasses¶

class monai.transforms.ConvertToMultiChannelBasedOnBratsClasses[source]

Convert labels to multi channels based on brats18 classes: label 1 is the necrotic and non-enhancing tumor core label 2 is the the peritumoral edema label 4 is the GD-enhancing tumor The possible classes are TC (Tumor core), WT (Whole tumor) and ET (Enhancing tumor).

__call__(img)[source]

data is an element which often comes from an iteration over an iterable, such as torch.utils.data.Dataset. This method should return an updated version of data. To simplify the input validations, most of the transforms assume that

• data is a Numpy ndarray, PyTorch Tensor or string

• the data shape can be:

1. string data without shape, LoadImage transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChannel expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirst expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

This method can optionally take additional arguments to help execute transformation operation.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

ndarray

Add extreme points of label to the image as a new channel. This transform generates extreme point from label and applies a gaussian filter. The pixel values in points image are rescaled to range [rescale_min, rescale_max] and added as a new channel to input image. The algorithm is described in Roth et al., Going to Extremes: Weakly Supervised Medical Image Segmentation https://arxiv.org/abs/2009.11988.

This transform only supports single channel labels (1, spatial_dim1, [spatial_dim2, …]). The background index is ignored when calculating extreme points.

Parameters
• background (int) – Class index of background label, defaults to 0.

• pert (float) – Random perturbation amount to add to the points, defaults to 0.0.

Raises
• ValueError – When no label image provided.

• ValueError – When label image is not single channel.

__call__(img, label=None, sigma=3.0, rescale_min=- 1.0, rescale_max=1.0)[source]
Parameters
• img (ndarray) – the image that we want to add new channel to.

• label (Optional[ndarray]) – label image to get extreme points from. Shape must be (1, spatial_dim1, [, spatial_dim2, …]). Doesn’t support one-hot labels.

• sigma (Union[Sequence[float], float, Sequence[Tensor], Tensor]) – if a list of values, must match the count of spatial dimensions of input data, and apply every value in the list to 1 spatial dimension. if only 1 value provided, use it for all spatial dimensions.

• rescale_min (float) – minimum value of output data.

• rescale_max (float) – maximum value of output data.

Return type

ndarray

randomize(label)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

#### TorchVision¶

class monai.transforms.TorchVision(name, *args, **kwargs)[source]

This is a wrapper transform for PyTorch TorchVision transform based on the specified transform name and args. As most of the TorchVision transforms only work for PIL image and PyTorch Tensor, this transform expects input data to be PyTorch Tensor, users can easily call ToTensor transform to convert a Numpy array to Tensor.

Parameters
• name (str) – The transform name in TorchVision package.

• args – parameters for the TorchVision transform.

• kwargs – parameters for the TorchVision transform.

__call__(img)[source]
Parameters

img (Tensor) – PyTorch Tensor data for the TorchVision transform.

## Dictionary Transforms¶

Dictionary-based wrapper of monai.transforms.SpatialPad. Performs padding to the data, symmetric for all sides or all on one side for each dimension.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• spatial_size (Union[Sequence[int], int]) – the spatial size of output data after padding. If its components have non-positive values, the corresponding size of input image will be used.

• method (Union[Method, str]) – {"symmetric", "end"} Pad image symmetric on every side or only pad at the end sides. Defaults to "symmetric".

• mode (Union[Sequence[Union[NumpyPadMode, str]], NumpyPadMode, str]) – {"constant", "edge", "linear_ramp", "maximum", "mean", "median", "minimum", "reflect", "symmetric", "wrap", "empty"} One of the listed string values or a user supplied function. Defaults to "constant". See also: https://numpy.org/doc/1.18/reference/generated/numpy.pad.html It also can be a sequence of string, each element corresponds to a key in keys.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, ndarray]

Returns

An updated dictionary version of data by applying the transform.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• spatial_border (Union[Sequence[int], int]) –

specified size for every spatial border. it can be 3 shapes:

• single int number, pad all the borders with the same size.

• length equals the length of image shape, pad every spatial dimension separately. for example, image shape(CHW) is [1, 4, 4], spatial_border is [2, 1], pad every border of H dim with 2, pad every border of W dim with 1, result shape is [1, 8, 6].

• length equals 2 x (length of image shape), pad every border of every dimension separately. for example, image shape(CHW) is [1, 4, 4], spatial_border is [1, 2, 3, 4], pad top of H dim with 1, pad bottom of H dim with 2, pad left of W dim with 3, pad right of W dim with 4. the result shape is [1, 7, 11].

• mode (Union[Sequence[Union[NumpyPadMode, str]], NumpyPadMode, str]) – {"constant", "edge", "linear_ramp", "maximum", "mean", "median", "minimum", "reflect", "symmetric", "wrap", "empty"} One of the listed string values or a user supplied function. Defaults to "constant". See also: https://numpy.org/doc/1.18/reference/generated/numpy.pad.html It also can be a sequence of string, each element corresponds to a key in keys.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, ndarray]

Returns

An updated dictionary version of data by applying the transform.

Pad the input data, so that the spatial sizes are divisible by k. Dictionary-based wrapper of monai.transforms.DivisiblePad.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• k (Union[Sequence[int], int]) – the target k for each spatial dimension. if k is negative or 0, the original size is preserved. if k is an int, the same k be applied to all the input spatial dimensions.

• mode (Union[Sequence[Union[NumpyPadMode, str]], NumpyPadMode, str]) – {"constant", "edge", "linear_ramp", "maximum", "mean", "median", "minimum", "reflect", "symmetric", "wrap", "empty"} One of the listed string values or a user supplied function. Defaults to "constant". See also: https://numpy.org/doc/1.18/reference/generated/numpy.pad.html It also can be a sequence of string, each element corresponds to a key in keys.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, ndarray]

Returns

An updated dictionary version of data by applying the transform.

#### SpatialCropd¶

class monai.transforms.SpatialCropd(keys, roi_center=None, roi_size=None, roi_start=None, roi_end=None)[source]

Dictionary-based wrapper of monai.transforms.SpatialCrop. Either a spatial center and size must be provided, or alternatively if center and size are not provided, the start and end coordinates of the ROI must be provided.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• roi_center (Optional[Sequence[int]]) – voxel coordinates for center of the crop ROI.

• roi_size (Optional[Sequence[int]]) – size of the crop ROI.

• roi_start (Optional[Sequence[int]]) – voxel coordinates for start of the crop ROI.

• roi_end (Optional[Sequence[int]]) – voxel coordinates for end of the crop ROI.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, ndarray]

Returns

An updated dictionary version of data by applying the transform.

#### CenterSpatialCropd¶

class monai.transforms.CenterSpatialCropd(keys, roi_size)[source]

Dictionary-based wrapper of monai.transforms.CenterSpatialCrop.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.MapTransform

• roi_size (Union[Sequence[int], int]) – the size of the crop region e.g. [224,224,128] If its components have non-positive values, the corresponding size of input image will be used.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, ndarray]

Returns

An updated dictionary version of data by applying the transform.

#### RandSpatialCropd¶

class monai.transforms.RandSpatialCropd(keys, roi_size, random_center=True, random_size=True)[source]

Dictionary-based version monai.transforms.RandSpatialCrop. Crop image with random size or specific size ROI. It can crop at a random position as center or at the image center. And allows to set the minimum size to limit the randomly generated ROI. Suppose all the expected fields specified by keys have same shape.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.MapTransform

• roi_size (Union[Sequence[int], int]) – if random_size is True, it specifies the minimum crop region. if random_size is False, it specifies the expected ROI size to crop. e.g. [224, 224, 128] If its components have non-positive values, the corresponding size of input image will be used.

• random_center (bool) – crop at random position as center or the image center.

• random_size (bool) – crop with random size or specific size ROI. The actual size is sampled from randint(roi_size, img_size).

__call__(data)[source]

Call self as a function.

Return type

Dict[Hashable, ndarray]

randomize(img_size)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

#### RandSpatialCropSamplesd¶

class monai.transforms.RandSpatialCropSamplesd(keys, roi_size, num_samples, random_center=True, random_size=True)[source]

Dictionary-based version monai.transforms.RandSpatialCropSamples. Crop image with random size or specific size ROI to generate a list of N samples. It can crop at a random position as center or at the image center. And allows to set the minimum size to limit the randomly generated ROI. Suppose all the expected fields specified by keys have same shape. It will return a list of dictionaries for all the cropped images.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.MapTransform

• roi_size (Union[Sequence[int], int]) – if random_size is True, the spatial size of the minimum crop region. if random_size is False, specify the expected ROI size to crop. e.g. [224, 224, 128]

• num_samples (int) – number of samples (crop regions) to take in the returned list.

• random_center (bool) – crop at random position as center or the image center.

• random_size (bool) – crop with random size or specific size ROI. The actual size is sampled from randint(roi_size, img_size).

Raises

ValueError – When num_samples is nonpositive.

__call__(data)[source]

Call self as a function.

Return type

List[Dict[Hashable, ndarray]]

randomize(data=None)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

set_random_state(seed=None, state=None)[source]

Set the random state locally, to control the randomness, the derived classes should use self.R instead of np.random to introduce random factors.

Parameters
• seed (Optional[int]) – set the random state with an integer seed.

• state (Optional[RandomState]) – set the random state with a np.random.RandomState object.

Raises

TypeError – When state is not an Optional[np.random.RandomState].

Return type

Randomizable

Returns

a Randomizable instance.

#### CropForegroundd¶

class monai.transforms.CropForegroundd(keys, source_key, select_fn=<function CropForegroundd.<lambda>>, channel_indices=None, margin=0, start_coord_key='foreground_start_coord', end_coord_key='foreground_end_coord')[source]

Dictionary-based version monai.transforms.CropForeground. Crop only the foreground object of the expected images. The typical usage is to help training and evaluation if the valid part is small in the whole medical image. The valid part can be determined by any field in the data with source_key, for example: - Select values > 0 in image field as the foreground and crop on all fields specified by keys. - Select label = 3 in label field as the foreground to crop on all fields specified by keys. - Select label > 0 in the third channel of a One-Hot label field as the foreground to crop all keys fields. Users can define arbitrary function to select expected foreground from the whole source image or specified channels. And it can also add margin to every dim of the bounding box of foreground object.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• source_key (str) – data source to generate the bounding box of foreground, can be image or label, etc.

• select_fn (Callable) – function to select expected foreground, default is to select values > 0.

• channel_indices (Union[Iterable[int], int, None]) – if defined, select foreground only on the specified channels of image. if None, select foreground on the whole image.

• margin (int) – add margin value to spatial dims of the bounding box, if only 1 value provided, use it for all dims.

• start_coord_key (str) – key to record the start coordinate of spatial bounding box for foreground.

• end_coord_key (str) – key to record the end coordinate of spatial bounding box for foreground.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, ndarray]

Returns

An updated dictionary version of data by applying the transform.

#### RandWeightedCropd¶

class monai.transforms.RandWeightedCropd(keys, w_key, spatial_size, num_samples=1, center_coord_key=None)[source]

Samples a list of num_samples image patches according to the provided weight_map.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• w_key (str) – key for the weight map. The corresponding value will be used as the sampling weights, it should be a single-channel array in size, for example, (1, spatial_dim_0, spatial_dim_1, …)

• spatial_size (Union[Sequence[int], int]) – the spatial size of the image patch e.g. [224, 224, 128]. If its components have non-positive values, the corresponding size of img will be used.

• num_samples (int) – number of samples (image patches) to take in the returned list.

• center_coord_key (Optional[str]) – if specified, the actual sampling location will be stored with the corresponding key.

__call__(data)[source]

Call self as a function.

Return type

List[Dict[Hashable, ndarray]]

randomize(weight_map)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

#### RandCropByPosNegLabeld¶

class monai.transforms.RandCropByPosNegLabeld(keys, label_key, spatial_size, pos=1.0, neg=1.0, num_samples=1, image_key=None, image_threshold=0.0, fg_indices_key=None, bg_indices_key=None)[source]

Dictionary-based version monai.transforms.RandCropByPosNegLabel. Crop random fixed sized regions with the center being a foreground or background voxel based on the Pos Neg Ratio. And will return a list of dictionaries for all the cropped images.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• label_key (str) – name of key for label image, this will be used for finding foreground/background.

• spatial_size (Union[Sequence[int], int]) – the spatial size of the crop region e.g. [224, 224, 128]. If its components have non-positive values, the corresponding size of data[label_key] will be used.

• pos (float) – used with neg together to calculate the ratio pos / (pos + neg) for the probability to pick a foreground voxel as a center rather than a background voxel.

• neg (float) – used with pos together to calculate the ratio pos / (pos + neg) for the probability to pick a foreground voxel as a center rather than a background voxel.

• num_samples (int) – number of samples (crop regions) to take in each list.

• image_key (Optional[str]) – if image_key is not None, use label == 0 & image > image_threshold to select the negative sample(background) center. so the crop center will only exist on valid image area.

• image_threshold (float) – if enabled image_key, use image > image_threshold to determine the valid image content area.

• fg_indices_key (Optional[str]) – if provided pre-computed foreground indices of label, will ignore above image_key and image_threshold, and randomly select crop centers based on them, need to provide fg_indices_key and bg_indices_key together, expect to be 1 dim array of spatial indices after flattening. a typical usage is to call FgBgToIndicesd transform first and cache the results.

• bg_indices_key (Optional[str]) – if provided pre-computed background indices of label, will ignore above image_key and image_threshold, and randomly select crop centers based on them, need to provide fg_indices_key and bg_indices_key together, expect to be 1 dim array of spatial indices after flattening. a typical usage is to call FgBgToIndicesd transform first and cache the results.

Raises
• ValueError – When pos or neg are negative.

• ValueError – When pos=0 and neg=0. Incompatible values.

__call__(data)[source]

Call self as a function.

Return type

List[Dict[Hashable, ndarray]]

randomize(label, fg_indices=None, bg_indices=None, image=None)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.MapTransform

• spatial_size (Union[Sequence[int], int]) – the spatial size of output data after padding or crop. If has non-positive values, the corresponding size of input image will be used (no padding).

• mode (Union[NumpyPadMode, str]) – {"constant", "edge", "linear_ramp", "maximum", "mean", "median", "minimum", "reflect", "symmetric", "wrap", "empty"} One of the listed string values or a user supplied function for padding. Defaults to "constant". See also: https://numpy.org/doc/1.18/reference/generated/numpy.pad.html

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, ndarray]

Returns

An updated dictionary version of data by applying the transform.

#### BoundingRectd¶

class monai.transforms.BoundingRectd(keys, bbox_key_postfix='bbox', select_fn=<function BoundingRectd.<lambda>>)[source]

Dictionary-based wrapper of monai.transforms.BoundingRect.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.MapTransform

• bbox_key_postfix (str) – the output bounding box coordinates will be written to the value of {key}_{bbox_key_postfix}.

• select_fn (Callable) – function to select expected foreground, default is to select values > 0.

__call__(data)[source]
Return type

Dict[Hashable, ndarray]

### Instensity (Dict)¶

#### RandGaussianNoised¶

class monai.transforms.RandGaussianNoised(keys, prob=0.1, mean=0.0, std=0.1)[source]

Dictionary-based version monai.transforms.RandGaussianNoise. Add Gaussian noise to image. This transform assumes all the expected fields have same shape.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• prob (float) – Probability to add Gaussian noise.

• mean (Union[Sequence[float], float]) – Mean or “centre” of the distribution.

• std (float) – Standard deviation (spread) of distribution.

__call__(data)[source]

Call self as a function.

Return type

Dict[Hashable, ndarray]

randomize(im_shape)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

#### ShiftIntensityd¶

class monai.transforms.ShiftIntensityd(keys, offset)[source]

Dictionary-based wrapper of monai.transforms.ShiftIntensity.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• offset (float) – offset value to shift the intensity of image.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, ndarray]

Returns

An updated dictionary version of data by applying the transform.

#### RandShiftIntensityd¶

class monai.transforms.RandShiftIntensityd(keys, offsets, prob=0.1)[source]

Dictionary-based version monai.transforms.RandShiftIntensity.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• offsets (Union[Tuple[float, float], float]) – offset range to randomly shift. if single number, offset value is picked from (-offsets, offsets).

• prob (float) – probability of rotating. (Default 0.1, with 10% probability it returns a rotated array.)

__call__(data)[source]

Call self as a function.

Return type

Dict[Hashable, ndarray]

randomize(data=None)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

#### ScaleIntensityd¶

class monai.transforms.ScaleIntensityd(keys, minv=0.0, maxv=1.0, factor=None)[source]

Dictionary-based wrapper of monai.transforms.ScaleIntensity. Scale the intensity of input image to the given value range (minv, maxv). If minv and maxv not provided, use factor to scale image by v = v * (1 + factor).

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• minv (float) – minimum value of output data.

• maxv (float) – maximum value of output data.

• factor (Optional[float]) – factor scale by v = v * (1 + factor).

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, ndarray]

Returns

An updated dictionary version of data by applying the transform.

#### RandScaleIntensityd¶

class monai.transforms.RandScaleIntensityd(keys, factors, prob=0.1)[source]

Dictionary-based version monai.transforms.RandScaleIntensity.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• factors (Union[Tuple[float, float], float]) – factor range to randomly scale by v = v * (1 + factor). if single number, factor value is picked from (-factors, factors).

• prob (float) – probability of rotating. (Default 0.1, with 10% probability it returns a rotated array.)

__call__(data)[source]

Call self as a function.

Return type

Dict[Hashable, ndarray]

randomize(data=None)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

#### NormalizeIntensityd¶

class monai.transforms.NormalizeIntensityd(keys, subtrahend=None, divisor=None, nonzero=False, channel_wise=False, dtype=<class 'numpy.float32'>)[source]

Dictionary-based wrapper of monai.transforms.NormalizeIntensity. This transform can normalize only non-zero values or entire image, and can also calculate mean and std on each channel separately.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.MapTransform

• subtrahend (Optional[ndarray]) – the amount to subtract by (usually the mean)

• divisor (Optional[ndarray]) – the amount to divide by (usually the standard deviation)

• nonzero (bool) – whether only normalize non-zero values.

• channel_wise (bool) – if using calculated mean and std, calculate on each channel separately or calculate on the entire image directly.

• dtype (dtype) – output data type, defaut to float32.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, ndarray]

Returns

An updated dictionary version of data by applying the transform.

#### ThresholdIntensityd¶

class monai.transforms.ThresholdIntensityd(keys, threshold, above=True, cval=0.0)[source]

Dictionary-based wrapper of monai.transforms.ThresholdIntensity.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.MapTransform

• threshold (float) – the threshold to filter intensity values.

• above (bool) – filter values above the threshold or below the threshold, default is True.

• cval (float) – value to fill the remaining parts of the image, default is 0.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, ndarray]

Returns

An updated dictionary version of data by applying the transform.

#### ScaleIntensityRanged¶

class monai.transforms.ScaleIntensityRanged(keys, a_min, a_max, b_min, b_max, clip=False)[source]

Dictionary-based wrapper of monai.transforms.ScaleIntensityRange.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.MapTransform

• a_min (float) – intensity original range min.

• a_max (float) – intensity original range max.

• b_min (float) – intensity target range min.

• b_max (float) – intensity target range max.

• clip (bool) – whether to perform clip after scaling.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, ndarray]

Returns

An updated dictionary version of data by applying the transform.

#### ScaleIntensityRangePercentilesd¶

class monai.transforms.ScaleIntensityRangePercentilesd(keys, lower, upper, b_min, b_max, clip=False, relative=False)[source]

Dictionary-based wrapper of monai.transforms.ScaleIntensityRangePercentiles.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.MapTransform

• lower (float) – lower percentile.

• upper (float) – upper percentile.

• b_min (float) – intensity target range min.

• b_max (float) – intensity target range max.

• clip (bool) – whether to perform clip after scaling.

• relative (bool) – whether to scale to the corresponding percentiles of [b_min, b_max]

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, ndarray]

Returns

An updated dictionary version of data by applying the transform.

Dictionary-based wrapper of monai.transforms.AdjustContrast. Changes image intensity by gamma. Each pixel/voxel intensity is updated as:

x = ((x - min) / intensity_range) ^ gamma * intensity_range + min

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.MapTransform

• gamma (float) – gamma value to adjust the contrast as function.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, ndarray]

Returns

An updated dictionary version of data by applying the transform.

Dictionary-based version monai.transforms.RandAdjustContrast. Randomly changes image intensity by gamma. Each pixel/voxel intensity is updated as:

x = ((x - min) / intensity_range) ^ gamma * intensity_range + min

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.MapTransform

• prob (float) – Probability of adjustment.

• gamma (Union[Tuple[float, float], float]) – Range of gamma values. If single number, value is picked from (0.5, gamma), default is (0.5, 4.5).

__call__(data)[source]

Call self as a function.

Return type

Dict[Hashable, ndarray]

randomize(data=None)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• mask_data (ndarray) – if mask data is single channel, apply to evey channel of input image. if multiple channels, the channel number must match input data. mask_data will be converted to bool values by mask_data > 0 before applying transform to input image.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, ndarray]

Returns

An updated dictionary version of data by applying the transform.

#### GaussianSmoothd¶

class monai.transforms.GaussianSmoothd(keys, sigma, approx='erf')[source]

Dictionary-based wrapper of monai.transforms.GaussianSmooth.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• sigma (Union[Sequence[float], float]) – if a list of values, must match the count of spatial dimensions of input data, and apply every value in the list to 1 spatial dimension. if only 1 value provided, use it for all spatial dimensions.

• approx (str) – discrete Gaussian kernel type, available options are “erf”, “sampled”, and “scalespace”. see also monai.networks.layers.GaussianFilter().

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, ndarray]

Returns

An updated dictionary version of data by applying the transform.

#### RandGaussianSmoothd¶

class monai.transforms.RandGaussianSmoothd(keys, sigma_x=(0.25, 1.5), sigma_y=(0.25, 1.5), sigma_z=(0.25, 1.5), approx='erf', prob=0.1)[source]

Dictionary-based wrapper of monai.transforms.GaussianSmooth.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• sigma_x (Tuple[float, float]) – randomly select sigma value for the first spatial dimension.

• sigma_y (Tuple[float, float]) – randomly select sigma value for the second spatial dimension if have.

• sigma_z (Tuple[float, float]) – randomly select sigma value for the third spatial dimension if have.

• approx (str) – discrete Gaussian kernel type, available options are “erf”, “sampled”, and “scalespace”. see also monai.networks.layers.GaussianFilter().

• prob (float) – probability of Gaussian smooth.

__call__(data)[source]

Call self as a function.

Return type

Dict[Hashable, ndarray]

randomize(data=None)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

#### GaussianSharpend¶

class monai.transforms.GaussianSharpend(keys, sigma1=3.0, sigma2=1.0, alpha=30.0, approx='erf')[source]

Dictionary-based wrapper of monai.transforms.GaussianSharpen.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• sigma1 (Union[Sequence[float], float]) – sigma parameter for the first gaussian kernel. if a list of values, must match the count of spatial dimensions of input data, and apply every value in the list to 1 spatial dimension. if only 1 value provided, use it for all spatial dimensions.

• sigma2 (Union[Sequence[float], float]) – sigma parameter for the second gaussian kernel. if a list of values, must match the count of spatial dimensions of input data, and apply every value in the list to 1 spatial dimension. if only 1 value provided, use it for all spatial dimensions.

• alpha (float) – weight parameter to compute the final result.

• approx (str) – discrete Gaussian kernel type, available options are “erf”, “sampled”, and “scalespace”. see also monai.networks.layers.GaussianFilter().

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, ndarray]

Returns

An updated dictionary version of data by applying the transform.

#### RandGaussianSharpend¶

class monai.transforms.RandGaussianSharpend(keys, sigma1_x=(0.5, 1.0), sigma1_y=(0.5, 1.0), sigma1_z=(0.5, 1.0), sigma2_x=0.5, sigma2_y=0.5, sigma2_z=0.5, alpha=(10.0, 30.0), approx='erf', prob=0.1)[source]

Dictionary-based wrapper of monai.transforms.GaussianSharpen.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• sigma1_x (Tuple[float, float]) – randomly select sigma value for the first spatial dimension of first gaussian kernel.

• sigma1_y (Tuple[float, float]) – randomly select sigma value for the second spatial dimension(if have) of first gaussian kernel.

• sigma1_z (Tuple[float, float]) – randomly select sigma value for the third spatial dimension(if have) of first gaussian kernel.

• sigma2_x (Union[Tuple[float, float], float]) – randomly select sigma value for the first spatial dimension of second gaussian kernel. if only 1 value X provided, it must be smaller than sigma1_x and randomly select from [X, sigma1_x].

• sigma2_y (Union[Tuple[float, float], float]) – randomly select sigma value for the second spatial dimension(if have) of second gaussian kernel. if only 1 value Y provided, it must be smaller than sigma1_y and randomly select from [Y, sigma1_y].

• sigma2_z (Union[Tuple[float, float], float]) – randomly select sigma value for the third spatial dimension(if have) of second gaussian kernel. if only 1 value Z provided, it must be smaller than sigma1_z and randomly select from [Z, sigma1_z].

• alpha (Tuple[float, float]) – randomly select weight parameter to compute the final result.

• approx (str) – discrete Gaussian kernel type, available options are “erf”, “sampled”, and “scalespace”. see also monai.networks.layers.GaussianFilter().

• prob (float) – probability of Gaussian sharpen.

__call__(data)[source]

Call self as a function.

randomize(data=None)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

#### RandHistogramShiftd¶

class monai.transforms.RandHistogramShiftd(keys, num_control_points=10, prob=0.1)[source]

Dictionary-based version monai.transforms.RandHistogramShift. Apply random nonlinear transform the the image’s intensity histogram.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.MapTransform

• num_control_points (Union[Tuple[int, int], int]) – number of control points governing the nonlinear intensity mapping. a smaller number of control points allows for larger intensity shifts. if two values provided, number of control points selecting from range (min_value, max_value).

• prob (float) – probability of histogram shift.

__call__(data)[source]

Call self as a function.

Return type

Dict[Hashable, ndarray]

randomize(data=None)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

### IO (Dict)¶

Dictionary-based wrapper of monai.transforms.LoadImage, must load image and metadata together. If loading a list of files in one key, stack them together and add a new dimension as the first dimension, and use the meta data of the first image to represent the stacked result. Note that the affine transform of all the stacked images should be same. The output metadata field will be created as key_{meta_key_postfix}.

It can automatically choose readers based on the supported suffixes and in below order: - User specified reader at runtime when call this loader. - Registered readers from the latest to the first in list. - Default readers: (nii, nii.gz -> NibabelReader), (png, jpg, bmp -> PILReader), (npz, npy -> NumpyReader), (others -> ITKReader).

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• dtype (Optional[dtype]) – if not None convert the loaded image data to this data type.

• meta_key_postfix (str) – use key_{postfix} to store the metadata of the nifti image, default is meta_dict. The meta data is a dictionary object. For example, load nifti file for image, store the metadata into image_meta_dict.

• overwriting (bool) – whether allow to overwrite existing meta data of same key. default is False, which will raise exception if encountering existing key.

Raises

KeyError – When not self.overwriting and key already exists in data.

Register a virtual subclass of an ABC.

Returns the subclass, to allow usage as a class decorator.

### Post-processing (Dict)¶

#### Activationsd¶

class monai.transforms.Activationsd(keys, sigmoid=False, softmax=False, other=None)[source]

Dictionary-based wrapper of monai.transforms.AddActivations. Add activation layers to the input data specified by keys.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to model output and label. See also: monai.transforms.compose.MapTransform

• sigmoid (Union[Sequence[bool], bool]) – whether to execute sigmoid function on model output before transform. it also can be a sequence of bool, each element corresponds to a key in keys.

• softmax (Union[Sequence[bool], bool]) – whether to execute softmax function on model output before transform. it also can be a sequence of bool, each element corresponds to a key in keys.

• other (Union[Sequence[Callable], Callable, None]) – callable function to execute other activation layers, for example: other = lambda x: torch.tanh(x). it also can be a sequence of Callable, each element corresponds to a key in keys.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, Tensor]

Returns

An updated dictionary version of data by applying the transform.

#### AsDiscreted¶

class monai.transforms.AsDiscreted(keys, argmax=False, to_onehot=False, n_classes=None, threshold_values=False, logit_thresh=0.5)[source]

Dictionary-based wrapper of monai.transforms.AsDiscrete.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to model output and label. See also: monai.transforms.compose.MapTransform

• argmax (Union[Sequence[bool], bool]) – whether to execute argmax function on input data before transform. it also can be a sequence of bool, each element corresponds to a key in keys.

• to_onehot (Union[Sequence[bool], bool]) – whether to convert input data into the one-hot format. Defaults to False. it also can be a sequence of bool, each element corresponds to a key in keys.

• n_classes (Union[Sequence[int], int, None]) – the number of classes to convert to One-Hot format. it also can be a sequence of int, each element corresponds to a key in keys.

• threshold_values (Union[Sequence[bool], bool]) – whether threshold the float value to int number 0 or 1, default is False. it also can be a sequence of bool, each element corresponds to a key in keys.

• logit_thresh (Union[Sequence[float], float]) – the threshold value for thresholding operation, default is 0.5. it also can be a sequence of float, each element corresponds to a key in keys.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, Tensor]

Returns

An updated dictionary version of data by applying the transform.

#### KeepLargestConnectedComponentd¶

class monai.transforms.KeepLargestConnectedComponentd(keys, applied_labels, independent=True, connectivity=None)[source]

Dictionary-based wrapper of monai.transforms.KeepLargestConnectedComponent.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• applied_labels (Union[Sequence[int], int]) – Labels for applying the connected component on. If only one channel. The pixel whose value is not in this list will remain unchanged. If the data is in one-hot format, this is the channel indices to apply transform.

• independent (bool) – consider several labels as a whole or independent, default is True. Example use case would be segment label 1 is liver and label 2 is liver tumor, in that case you want this “independent” to be specified as False.

• connectivity (Optional[int]) – Maximum number of orthogonal hops to consider a pixel/voxel as a neighbor. Accepted values are ranging from 1 to input.ndim. If None, a full connectivity of input.ndim is used.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, Tensor]

Returns

An updated dictionary version of data by applying the transform.

#### LabelToContourd¶

class monai.transforms.LabelToContourd(keys, kernel_type='Laplace')[source]

Dictionary-based wrapper of monai.transforms.LabelToContour.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• kernel_type (str) – the method applied to do edge detection, default is “Laplace”.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, Tensor]

Returns

An updated dictionary version of data by applying the transform.

#### Ensembled¶

class monai.transforms.Ensembled(keys, ensemble, output_key=None)[source]

Base class of dictionary-based ensemble transforms.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be stack and execute ensemble. if only 1 key provided, suppose it’s a PyTorch Tensor with data stacked on dimension E.

• output_key (Optional[str]) – the key to store ensemble result in the dictionary.

• ensemble (Callable[[Union[Sequence[Tensor], Tensor]], Tensor]) – callable method to execute ensemble on specified data. if only 1 key provided in keys, output_key can be None and use keys as default.

Raises
• TypeError – When ensemble is not callable.

• ValueError – When len(keys) > 1 and output_key=None. Incompatible values.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, Tensor]

Returns

An updated dictionary version of data by applying the transform.

#### MeanEnsembled¶

class monai.transforms.MeanEnsembled(keys, output_key=None, weights=None)[source]

Dictionary-based wrapper of monai.transforms.MeanEnsemble.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be stack and execute ensemble. if only 1 key provided, suppose it’s a PyTorch Tensor with data stacked on dimension E.

• output_key (Optional[str]) – the key to store ensemble result in the dictionary. if only 1 key provided in keys, output_key can be None and use keys as default.

• weights (Union[Sequence[float], Tensor, ndarray, None]) – can be a list or tuple of numbers for input data with shape: [E, B, C, H, W[, D]]. or a Numpy ndarray or a PyTorch Tensor data. the weights will be added to input data from highest dimension, for example: 1. if the weights only has 1 dimension, it will be added to the E dimension of input data. 2. if the weights has 3 dimensions, it will be added to E, B and C dimensions. it’s a typical practice to add weights for different classes: to ensemble 3 segmentation model outputs, every output has 4 channels(classes), so the input data shape can be: [3, B, 4, H, W, D]. and add different weights for different classes, so the weights shape can be: [3, 1, 4]. for example: weights = [[[1, 2, 3, 4]], [[4, 3, 2, 1]], [[1, 1, 1, 1]]].

#### VoteEnsembled¶

class monai.transforms.VoteEnsembled(keys, output_key=None, num_classes=None)[source]

Dictionary-based wrapper of monai.transforms.VoteEnsemble.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be stack and execute ensemble. if only 1 key provided, suppose it’s a PyTorch Tensor with data stacked on dimension E.

• output_key (Optional[str]) – the key to store ensemble result in the dictionary. if only 1 key provided in keys, output_key can be None and use keys as default.

• num_classes (Optional[int]) – if the input is single channel data instead of One-Hot, we can’t get class number from channel, need to explicitly specify the number of classes to vote.

### Spatial (Dict)¶

#### Spacingd¶

class monai.transforms.Spacingd(keys, pixdim, diagonal=False, mode=<GridSampleMode.BILINEAR: 'bilinear'>, padding_mode=<GridSamplePadMode.BORDER: 'border'>, align_corners=False, dtype=<class 'numpy.float64'>, meta_key_postfix='meta_dict')[source]

Dictionary-based wrapper of monai.transforms.Spacing.

This transform assumes the data dictionary has a key for the input data’s metadata and contains affine field. The key is formed by key_{meta_key_postfix}.

After resampling the input array, this transform will write the new affine to the affine field of metadata which is formed by key_{meta_key_postfix}.

Parameters
• pixdim (Sequence[float]) – output voxel spacing.

• diagonal (bool) –

whether to resample the input to have a diagonal affine matrix. If True, the input data is resampled to the following affine:

np.diag((pixdim_0, pixdim_1, pixdim_2, 1))

This effectively resets the volume to the world coordinate system (RAS+ in nibabel). The original orientation, rotation, shearing are not preserved.

If False, the axes orientation, orthogonal rotation and translations components from the original affine will be preserved in the target affine. This option will not flip/swap axes against the original ones.

• mode (Union[Sequence[Union[GridSampleMode, str]], GridSampleMode, str]) – {"bilinear", "nearest"} Interpolation mode to calculate output values. Defaults to "bilinear". See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample It also can be a sequence of string, each element corresponds to a key in keys.

• padding_mode (Union[Sequence[Union[GridSamplePadMode, str]], GridSamplePadMode, str]) – {"zeros", "border", "reflection"} Padding mode for outside grid values. Defaults to "border". See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample It also can be a sequence of string, each element corresponds to a key in keys.

• align_corners (Union[Sequence[bool], bool]) – Geometrically, we consider the pixels of the input as squares rather than points. See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample It also can be a sequence of bool, each element corresponds to a key in keys.

• dtype (Union[Sequence[dtype], dtype, None]) – data type for resampling computation. Defaults to np.float64 for best precision. If None, use the data type of input data. To be compatible with other modules, the output data type is always np.float32. It also can be a sequence of np.dtype, each element corresponds to a key in keys.

• meta_key_postfix (str) – use key_{postfix} to to fetch the meta data according to the key data, default is meta_dict, the meta data is a dictionary object. For example, to handle key image, read/write affine matrices from the metadata image_meta_dict dictionary’s affine field.

Raises

TypeError – When meta_key_postfix is not a str.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Union[Hashable, str], Union[ndarray, Dict[str, ndarray]]]

Returns

An updated dictionary version of data by applying the transform.

#### Orientationd¶

class monai.transforms.Orientationd(keys, axcodes=None, as_closest_canonical=False, labels=(('L', 'R'), ('P', 'A'), ('I', 'S')), meta_key_postfix='meta_dict')[source]

Dictionary-based wrapper of monai.transforms.Orientation.

This transform assumes the data dictionary has a key for the input data’s metadata and contains affine field. The key is formed by key_{meta_key_postfix}.

After reorienting the input array, this transform will write the new affine to the affine field of metadata which is formed by key_{meta_key_postfix}.

Parameters
• axcodes (Optional[str]) – N elements sequence for spatial ND input’s orientation. e.g. axcodes=’RAS’ represents 3D orientation: (Left, Right), (Posterior, Anterior), (Inferior, Superior). default orientation labels options are: ‘L’ and ‘R’ for the first dimension, ‘P’ and ‘A’ for the second, ‘I’ and ‘S’ for the third.

• as_closest_canonical (bool) – if True, load the image as closest to canonical axis format.

• labels (Optional[Sequence[Tuple[str, str]]]) – optional, None or sequence of (2,) sequences (2,) sequences are labels for (beginning, end) of output axis. Defaults to (('L', 'R'), ('P', 'A'), ('I', 'S')).

• meta_key_postfix (str) – use key_{postfix} to to fetch the meta data according to the key data, default is meta_dict, the meta data is a dictionary object. For example, to handle key image, read/write affine matrices from the metadata image_meta_dict dictionary’s affine field.

Raises

TypeError – When meta_key_postfix is not a str.

nibabel.orientations.ornt2axcodes.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Union[Hashable, str], Union[ndarray, Dict[str, ndarray]]]

Returns

An updated dictionary version of data by applying the transform.

#### Flipd¶

class monai.transforms.Flipd(keys, spatial_axis=None)[source]

Dictionary-based wrapper of monai.transforms.Flip.

See numpy.flip for additional details. https://docs.scipy.org/doc/numpy/reference/generated/numpy.flip.html

Parameters
• keys (Union[Collection[Hashable], Hashable]) – Keys to pick data for transformation.

• spatial_axis (Union[Sequence[int], int, None]) – Spatial axes along which to flip over. Default is None.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, ndarray]

Returns

An updated dictionary version of data by applying the transform.

#### RandFlipd¶

class monai.transforms.RandFlipd(keys, prob=0.1, spatial_axis=None)[source]

Dictionary-based version monai.transforms.RandFlip.

See numpy.flip for additional details. https://docs.scipy.org/doc/numpy/reference/generated/numpy.flip.html

Parameters
• keys (Union[Collection[Hashable], Hashable]) – Keys to pick data for transformation.

• prob (float) – Probability of flipping.

• spatial_axis (Union[Sequence[int], int, None]) – Spatial axes along which to flip over. Default is None.

__call__(data)[source]

Call self as a function.

Return type

Dict[Hashable, ndarray]

randomize(data=None)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

#### Rotated¶

Dictionary-based wrapper of monai.transforms.Rotate.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – Keys to pick data for transformation.

• angle (Union[Sequence[float], float]) – Rotation angle(s) in radians.

• keep_size (bool) – If it is False, the output shape is adapted so that the input array is contained completely in the output. If it is True, the output shape is the same as the input. Default is True.

• mode (Union[Sequence[Union[GridSampleMode, str]], GridSampleMode, str]) – {"bilinear", "nearest"} Interpolation mode to calculate output values. Defaults to "bilinear". See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample It also can be a sequence of string, each element corresponds to a key in keys.

• padding_mode (Union[Sequence[Union[GridSamplePadMode, str]], GridSamplePadMode, str]) – {"zeros", "border", "reflection"} Padding mode for outside grid values. Defaults to "border". See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample It also can be a sequence of string, each element corresponds to a key in keys.

• align_corners (Union[Sequence[bool], bool]) – Defaults to False. See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample It also can be a sequence of bool, each element corresponds to a key in keys.

• dtype (Union[Sequence[Optional[dtype]], dtype, None]) – data type for resampling computation. Defaults to np.float64 for best precision. If None, use the data type of input data. To be compatible with other modules, the output data type is always np.float32. It also can be a sequence of dtype or None, each element corresponds to a key in keys.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, ndarray]

Returns

An updated dictionary version of data by applying the transform.

#### RandRotated¶

class monai.transforms.RandRotated(keys, range_x=0.0, range_y=0.0, range_z=0.0, prob=0.1, keep_size=True, mode=<GridSampleMode.BILINEAR: 'bilinear'>, padding_mode=<GridSamplePadMode.BORDER: 'border'>, align_corners=False, dtype=<class 'numpy.float64'>)[source]

Dictionary-based version monai.transforms.RandRotate Randomly rotates the input arrays.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – Keys to pick data for transformation.

• range_x (Union[Tuple[float, float], float]) – Range of rotation angle in radians in the plane defined by the first and second axes. If single number, angle is uniformly sampled from (-range_x, range_x).

• range_y (Union[Tuple[float, float], float]) – Range of rotation angle in radians in the plane defined by the first and third axes. If single number, angle is uniformly sampled from (-range_y, range_y).

• range_z (Union[Tuple[float, float], float]) – Range of rotation angle in radians in the plane defined by the second and third axes. If single number, angle is uniformly sampled from (-range_z, range_z).

• prob (float) – Probability of rotation.

• keep_size (bool) – If it is False, the output shape is adapted so that the input array is contained completely in the output. If it is True, the output shape is the same as the input. Default is True.

• mode (Union[Sequence[Union[GridSampleMode, str]], GridSampleMode, str]) – {"bilinear", "nearest"} Interpolation mode to calculate output values. Defaults to "bilinear". See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample It also can be a sequence of string, each element corresponds to a key in keys.

• padding_mode (Union[Sequence[Union[GridSamplePadMode, str]], GridSamplePadMode, str]) – {"zeros", "border", "reflection"} Padding mode for outside grid values. Defaults to "border". See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample It also can be a sequence of string, each element corresponds to a key in keys.

• align_corners (Union[Sequence[bool], bool]) – Defaults to False. See also: https://pytorch.org/docs/stable/nn.functional.html#interpolate It also can be a sequence of bool, each element corresponds to a key in keys.

• dtype (Union[Sequence[Optional[dtype]], dtype, None]) – data type for resampling computation. Defaults to np.float64 for best precision. If None, use the data type of input data. To be compatible with other modules, the output data type is always np.float32. It also can be a sequence of dtype or None, each element corresponds to a key in keys.

__call__(data)[source]

Call self as a function.

Return type

Dict[Hashable, ndarray]

randomize(data=None)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

#### Zoomd¶

Dictionary-based wrapper of monai.transforms.Zoom.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – Keys to pick data for transformation.

• zoom (Union[Sequence[float], float]) – The zoom factor along the spatial axes. If a float, zoom is the same for each spatial axis. If a sequence, zoom should contain one value for each spatial axis.

• mode (Union[Sequence[Union[InterpolateMode, str]], InterpolateMode, str]) – {"nearest", "linear", "bilinear", "bicubic", "trilinear", "area"} The interpolation mode. Defaults to "area". See also: https://pytorch.org/docs/stable/nn.functional.html#interpolate It also can be a sequence of string, each element corresponds to a key in keys.

• padding_mode (Union[Sequence[Union[NumpyPadMode, str]], NumpyPadMode, str]) – {"constant", "edge”, "linear_ramp”, "maximum”, "mean”, “median”, "minimum”, “reflect”, "symmetric”, "wrap”, "empty”, "<function>”} The mode to pad data after zooming. See also: https://numpy.org/doc/stable/reference/generated/numpy.pad.html

• align_corners (Union[Sequence[Optional[bool]], bool, None]) – This only has an effect when mode is ‘linear’, ‘bilinear’, ‘bicubic’ or ‘trilinear’. Default: None. See also: https://pytorch.org/docs/stable/nn.functional.html#interpolate It also can be a sequence of bool or None, each element corresponds to a key in keys.

• keep_size (bool) – Should keep original size (pad if needed), default is True.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, ndarray]

Returns

An updated dictionary version of data by applying the transform.

#### RandZoomd¶

Dict-based version monai.transforms.RandZoom.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – Keys to pick data for transformation.

• prob (float) – Probability of zooming.

• min_zoom (Union[Sequence[float], float]) – Min zoom factor. Can be float or sequence same size as image. If a float, select a random factor from [min_zoom, max_zoom] then apply to all spatial dims to keep the original spatial shape ratio. If a sequence, min_zoom should contain one value for each spatial axis. If 2 values provided for 3D data, use the first value for both H & W dims to keep the same zoom ratio.

• max_zoom (Union[Sequence[float], float]) – Max zoom factor. Can be float or sequence same size as image. If a float, select a random factor from [min_zoom, max_zoom] then apply to all spatial dims to keep the original spatial shape ratio. If a sequence, max_zoom should contain one value for each spatial axis. If 2 values provided for 3D data, use the first value for both H & W dims to keep the same zoom ratio.

• mode (Union[Sequence[Union[InterpolateMode, str]], InterpolateMode, str]) – {"nearest", "linear", "bilinear", "bicubic", "trilinear", "area"} The interpolation mode. Defaults to "area". See also: https://pytorch.org/docs/stable/nn.functional.html#interpolate It also can be a sequence of string, each element corresponds to a key in keys.

• padding_mode (Union[Sequence[Union[NumpyPadMode, str]], NumpyPadMode, str]) – {"constant", "edge”, "linear_ramp”, "maximum”, "mean”, “median”, "minimum”, “reflect”, "symmetric”, "wrap”, "empty”, "<function>”} The mode to pad data after zooming. See also: https://numpy.org/doc/stable/reference/generated/numpy.pad.html

• align_corners (Union[Sequence[Optional[bool]], bool, None]) – This only has an effect when mode is ‘linear’, ‘bilinear’, ‘bicubic’ or ‘trilinear’. Default: None. See also: https://pytorch.org/docs/stable/nn.functional.html#interpolate It also can be a sequence of bool or None, each element corresponds to a key in keys.

• keep_size (bool) – Should keep original size (pad if needed), default is True.

__call__(data)[source]

Call self as a function.

Return type

Dict[Hashable, ndarray]

randomize(data=None)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

#### RandRotate90d¶

class monai.transforms.RandRotate90d(keys, prob=0.1, max_k=3, spatial_axes=(0, 1))[source]

Dictionary-based version monai.transforms.RandRotate90. With probability prob, input arrays are rotated by 90 degrees in the plane specified by spatial_axes.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• prob (float) – probability of rotating. (Default 0.1, with 10% probability it returns a rotated array.)

• max_k (int) – number of rotations will be sampled from np.random.randint(max_k) + 1. (Default 3)

• spatial_axes (Tuple[int, int]) – 2 int numbers, defines the plane to rotate with 2 spatial axes. Default: (0, 1), this is the first two axis in spatial dimensions.

__call__(data)[source]

Call self as a function.

Return type

Mapping[Hashable, ndarray]

randomize(data=None)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

#### Rotate90d¶

class monai.transforms.Rotate90d(keys, k=1, spatial_axes=(0, 1))[source]

Dictionary-based wrapper of monai.transforms.Rotate90.

Parameters
• k (int) – number of times to rotate by 90 degrees.

• spatial_axes (Tuple[int, int]) – 2 int numbers, defines the plane to rotate with 2 spatial axes. Default: (0, 1), this is the first two axis in spatial dimensions.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, ndarray]

Returns

An updated dictionary version of data by applying the transform.

#### Resized¶

class monai.transforms.Resized(keys, spatial_size, mode=<InterpolateMode.AREA: 'area'>, align_corners=None)[source]

Dictionary-based wrapper of monai.transforms.Resize.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• spatial_size (Union[Sequence[int], int]) – expected shape of spatial dimensions after resize operation. if the components of the spatial_size are non-positive values, the transform will use the corresponding components of img size. For example, spatial_size=(32, -1) will be adapted to (32, 64) if the second spatial dimension size of img is 64.

• mode (Union[Sequence[Union[InterpolateMode, str]], InterpolateMode, str]) – {"nearest", "linear", "bilinear", "bicubic", "trilinear", "area"} The interpolation mode. Defaults to "area". See also: https://pytorch.org/docs/stable/nn.functional.html#interpolate It also can be a sequence of string, each element corresponds to a key in keys.

• align_corners (Union[Sequence[Optional[bool]], bool, None]) – This only has an effect when mode is ‘linear’, ‘bilinear’, ‘bicubic’ or ‘trilinear’. Default: None. See also: https://pytorch.org/docs/stable/nn.functional.html#interpolate It also can be a sequence of bool or None, each element corresponds to a key in keys.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, ndarray]

Returns

An updated dictionary version of data by applying the transform.

#### RandAffined¶

class monai.transforms.RandAffined(keys, spatial_size=None, prob=0.1, rotate_range=None, shear_range=None, translate_range=None, scale_range=None, mode=<GridSampleMode.BILINEAR: 'bilinear'>, padding_mode=<GridSamplePadMode.REFLECTION: 'reflection'>, as_tensor_output=True, device=None)[source]

Dictionary-based wrapper of monai.transforms.RandAffine.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed.

• spatial_size (Union[Sequence[int], int, None]) – output image spatial size. if spatial_size and self.spatial_size are not defined, or smaller than 1, the transform will use the spatial size of img. if the components of the spatial_size are non-positive values, the transform will use the corresponding components of img size. For example, spatial_size=(32, -1) will be adapted to (32, 64) if the second spatial dimension size of img is 64.

• prob (float) – probability of returning a randomized affine grid. defaults to 0.1, with 10% chance returns a randomized grid.

• rotate_range (Union[Sequence[float], float, None]) – angle range in radians. rotate_range[0] with be used to generate the 1st rotation parameter from uniform[-rotate_range[0], rotate_range[0]). Similarly, rotate_range[1] and rotate_range[2] are used in 3D affine for the range of 2nd and 3rd axes.

• shear_range (Union[Sequence[float], float, None]) – shear_range[0] with be used to generate the 1st shearing parameter from uniform[-shear_range[0], shear_range[0]). Similarly, shear_range[1] to shear_range[N] controls the range of the uniform distribution used to generate the 2nd to N-th parameter.

• translate_range (Union[Sequence[float], float, None]) – translate_range[0] with be used to generate the 1st shift parameter from uniform[-translate_range[0], translate_range[0]). Similarly, translate_range[1] to translate_range[N] controls the range of the uniform distribution used to generate the 2nd to N-th parameter.

• scale_range (Union[Sequence[float], float, None]) – scaling_range[0] with be used to generate the 1st scaling factor from uniform[-scale_range[0], scale_range[0]) + 1.0. Similarly, scale_range[1] to scale_range[N] controls the range of the uniform distribution used to generate the 2nd to N-th parameter.

• mode (Union[Sequence[Union[GridSampleMode, str]], GridSampleMode, str]) – {"bilinear", "nearest"} Interpolation mode to calculate output values. Defaults to "bilinear". See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample It also can be a sequence of string, each element corresponds to a key in keys.

• padding_mode (Union[Sequence[Union[GridSamplePadMode, str]], GridSamplePadMode, str]) – {"zeros", "border", "reflection"} Padding mode for outside grid values. Defaults to "reflection". See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample It also can be a sequence of string, each element corresponds to a key in keys.

• as_tensor_output (bool) – the computation is implemented using pytorch tensors, this option specifies whether to convert it back to numpy arrays.

• device (Optional[device]) – device on which the tensor will be allocated.

• monai.transforms.compose.MapTransform

• RandAffineGrid for the random affine parameters configurations.

__call__(data)[source]

Call self as a function.

Return type

Dict[Hashable, Union[ndarray, Tensor]]

randomize(data=None)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

set_random_state(seed=None, state=None)[source]

Set the random state locally, to control the randomness, the derived classes should use self.R instead of np.random to introduce random factors.

Parameters
• seed (Optional[int]) – set the random state with an integer seed.

• state (Optional[RandomState]) – set the random state with a np.random.RandomState object.

Raises

TypeError – When state is not an Optional[np.random.RandomState].

Return type

RandAffined

Returns

a Randomizable instance.

#### Rand2DElasticd¶

class monai.transforms.Rand2DElasticd(keys, spacing, magnitude_range, spatial_size=None, prob=0.1, rotate_range=None, shear_range=None, translate_range=None, scale_range=None, mode=<GridSampleMode.BILINEAR: 'bilinear'>, padding_mode=<GridSamplePadMode.REFLECTION: 'reflection'>, as_tensor_output=False, device=None)[source]

Dictionary-based wrapper of monai.transforms.Rand2DElastic.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed.

• spacing (Union[Tuple[float, float], float]) – distance in between the control points.

• magnitude_range (Tuple[float, float]) – 2 int numbers, the random offsets will be generated from uniform[magnitude[0], magnitude[1]).

• spatial_size (Union[Sequence[int], int, None]) – specifying output image spatial size [h, w]. if spatial_size and self.spatial_size are not defined, or smaller than 1, the transform will use the spatial size of img. if the components of the spatial_size are non-positive values, the transform will use the corresponding components of img size. For example, spatial_size=(32, -1) will be adapted to (32, 64) if the second spatial dimension size of img is 64.

• prob (float) – probability of returning a randomized affine grid. defaults to 0.1, with 10% chance returns a randomized grid, otherwise returns a spatial_size centered area extracted from the input image.

• rotate_range (Union[Sequence[float], float, None]) – angle range in radians. rotate_range[0] with be used to generate the 1st rotation parameter from uniform[-rotate_range[0], rotate_range[0]).

• shear_range (Union[Sequence[float], float, None]) – shear_range[0] with be used to generate the 1st shearing parameter from uniform[-shear_range[0], shear_range[0]). Similarly, shear_range[1] controls the range of the uniform distribution used to generate the 2nd parameter.

• translate_range (Union[Sequence[float], float, None]) – translate_range[0] with be used to generate the 1st shift parameter from uniform[-translate_range[0], translate_range[0]). Similarly, translate_range[1] controls the range of the uniform distribution used to generate the 2nd parameter.

• scale_range (Union[Sequence[float], float, None]) – scaling_range[0] with be used to generate the 1st scaling factor from uniform[-scale_range[0], scale_range[0]) + 1.0. Similarly, scale_range[1] controls the range of the uniform distribution used to generate the 2nd parameter.

• mode (Union[Sequence[Union[GridSampleMode, str]], GridSampleMode, str]) – {"bilinear", "nearest"} Interpolation mode to calculate output values. Defaults to "bilinear". See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample It also can be a sequence of string, each element corresponds to a key in keys.

• padding_mode (Union[Sequence[Union[GridSamplePadMode, str]], GridSamplePadMode, str]) – {"zeros", "border", "reflection"} Padding mode for outside grid values. Defaults to "reflection". See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample It also can be a sequence of string, each element corresponds to a key in keys.

• as_tensor_output (bool) – the computation is implemented using pytorch tensors, this option specifies whether to convert it back to numpy arrays.

• device (Optional[device]) – device on which the tensor will be allocated.

• RandAffineGrid for the random affine parameters configurations.

• Affine for the affine transformation parameters configurations.

__call__(data)[source]

Call self as a function.

Return type

Dict[Hashable, Union[ndarray, Tensor]]

randomize(spatial_size)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

set_random_state(seed=None, state=None)[source]

Set the random state locally, to control the randomness, the derived classes should use self.R instead of np.random to introduce random factors.

Parameters
• seed (Optional[int]) – set the random state with an integer seed.

• state (Optional[RandomState]) – set the random state with a np.random.RandomState object.

Raises

TypeError – When state is not an Optional[np.random.RandomState].

Return type

Rand2DElasticd

Returns

a Randomizable instance.

#### Rand3DElasticd¶

class monai.transforms.Rand3DElasticd(keys, sigma_range, magnitude_range, spatial_size=None, prob=0.1, rotate_range=None, shear_range=None, translate_range=None, scale_range=None, mode=<GridSampleMode.BILINEAR: 'bilinear'>, padding_mode=<GridSamplePadMode.REFLECTION: 'reflection'>, as_tensor_output=False, device=None)[source]

Dictionary-based wrapper of monai.transforms.Rand3DElastic.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed.

• sigma_range (Tuple[float, float]) – a Gaussian kernel with standard deviation sampled from uniform[sigma_range[0], sigma_range[1]) will be used to smooth the random offset grid.

• magnitude_range (Tuple[float, float]) – the random offsets on the grid will be generated from uniform[magnitude[0], magnitude[1]).

• spatial_size (Union[Sequence[int], int, None]) – specifying output image spatial size [h, w, d]. if spatial_size and self.spatial_size are not defined, or smaller than 1, the transform will use the spatial size of img. if the components of the spatial_size are non-positive values, the transform will use the corresponding components of img size. For example, spatial_size=(32, 32, -1) will be adapted to (32, 32, 64) if the third spatial dimension size of img is 64.

• prob (float) – probability of returning a randomized affine grid. defaults to 0.1, with 10% chance returns a randomized grid, otherwise returns a spatial_size centered area extracted from the input image.

• rotate_range (Union[Sequence[float], float, None]) – angle range in radians. rotate_range[0] with be used to generate the 1st rotation parameter from uniform[-rotate_range[0], rotate_range[0]). Similarly, rotate_range[1] and rotate_range[2] are used in 3D affine for the range of 2nd and 3rd axes.

• shear_range (Union[Sequence[float], float, None]) – shear_range[0] with be used to generate the 1st shearing parameter from uniform[-shear_range[0], shear_range[0]). Similarly, shear_range[1] and shear_range[2] controls the range of the uniform distribution used to generate the 2nd and 3rd parameters.

• translate_range (Union[Sequence[float], float, None]) – translate_range[0] with be used to generate the 1st shift parameter from uniform[-translate_range[0], translate_range[0]). Similarly, translate_range[1] and translate_range[2] controls the range of the uniform distribution used to generate the 2nd and 3rd parameters.

• scale_range (Union[Sequence[float], float, None]) – scaling_range[0] with be used to generate the 1st scaling factor from uniform[-scale_range[0], scale_range[0]) + 1.0. Similarly, scale_range[1] and scale_range[2] controls the range of the uniform distribution used to generate the 2nd and 3rd parameters.

• mode (Union[Sequence[Union[GridSampleMode, str]], GridSampleMode, str]) – {"bilinear", "nearest"} Interpolation mode to calculate output values. Defaults to "bilinear". See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample It also can be a sequence of string, each element corresponds to a key in keys.

• padding_mode (Union[Sequence[Union[GridSamplePadMode, str]], GridSamplePadMode, str]) – {"zeros", "border", "reflection"} Padding mode for outside grid values. Defaults to "reflection". See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample It also can be a sequence of string, each element corresponds to a key in keys.

• as_tensor_output (bool) – the computation is implemented using pytorch tensors, this option specifies whether to convert it back to numpy arrays.

• device (Optional[device]) – device on which the tensor will be allocated.

• RandAffineGrid for the random affine parameters configurations.

• Affine for the affine transformation parameters configurations.

__call__(data)[source]

Call self as a function.

Return type

Dict[Hashable, Union[ndarray, Tensor]]

randomize(grid_size)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

set_random_state(seed=None, state=None)[source]

Set the random state locally, to control the randomness, the derived classes should use self.R instead of np.random to introduce random factors.

Parameters
• seed (Optional[int]) – set the random state with an integer seed.

• state (Optional[RandomState]) – set the random state with a np.random.RandomState object.

Raises

TypeError – When state is not an Optional[np.random.RandomState].

Return type

Rand3DElasticd

Returns

a Randomizable instance.

### Utility (Dict)¶

#### Identityd¶

class monai.transforms.Identityd(keys)[source]

Dictionary-based wrapper of monai.transforms.Identity.

Parameters

keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, ndarray]

Returns

An updated dictionary version of data by applying the transform.

#### AsChannelFirstd¶

class monai.transforms.AsChannelFirstd(keys, channel_dim=- 1)[source]

Dictionary-based wrapper of monai.transforms.AsChannelFirst.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• channel_dim (int) – which dimension of input image is the channel, default is the last dimension.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, ndarray]

Returns

An updated dictionary version of data by applying the transform.

#### AsChannelLastd¶

class monai.transforms.AsChannelLastd(keys, channel_dim=0)[source]

Dictionary-based wrapper of monai.transforms.AsChannelLast.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• channel_dim (int) – which dimension of input image is the channel, default is the first dimension.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, ndarray]

Returns

An updated dictionary version of data by applying the transform.

Parameters

keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, Union[ndarray, Tensor]]

Returns

An updated dictionary version of data by applying the transform.

#### RepeatChanneld¶

class monai.transforms.RepeatChanneld(keys, repeats)[source]

Dictionary-based wrapper of monai.transforms.RepeatChannel.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• repeats (int) – the number of repetitions for each element.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, ndarray]

Returns

An updated dictionary version of data by applying the transform.

#### SplitChanneld¶

class monai.transforms.SplitChanneld(keys, output_postfixes=None, channel_dim=None)[source]

Dictionary-based wrapper of monai.transforms.SplitChannel. All the input specified by keys should be split into same count of data.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• output_postfixes (Optional[Sequence[str]]) – the postfixes to construct keys to store split data. for example: if the key of input data is pred and split 2 classes, the output data keys will be: pred_(output_postfixes[0]), pred_(output_postfixes[1]) if None, using the index number: pred_0, pred_1, … pred_N.

• channel_dim (Optional[int]) – which dimension of input image is the channel, default to None to automatically select: if data is numpy array, channel_dim is 0 as numpy array is used in the pre transforms, if PyTorch Tensor, channel_dim is 1 as in most of the cases Tensor is uses in the post transforms.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, Union[ndarray, Tensor]]

Returns

An updated dictionary version of data by applying the transform.

#### CastToTyped¶

class monai.transforms.CastToTyped(keys, dtype=<class 'numpy.float32'>)[source]

Dictionary-based wrapper of monai.transforms.CastToType.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• dtype (Union[Sequence[Union[dtype, dtype]], dtype, dtype]) – convert image to this data type, default is np.float32. it also can be a sequence of np.dtype or torch.dtype, each element corresponds to a key in keys.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, Union[ndarray, Tensor]]

Returns

An updated dictionary version of data by applying the transform.

#### ToTensord¶

class monai.transforms.ToTensord(keys)[source]

Dictionary-based wrapper of monai.transforms.ToTensor.

Parameters

keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, Tensor]

Returns

An updated dictionary version of data by applying the transform.

#### ToNumpyd¶

class monai.transforms.ToNumpyd(keys)[source]

Dictionary-based wrapper of monai.transforms.ToNumpy.

Parameters

keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, ndarray]

Returns

An updated dictionary version of data by applying the transform.

#### DeleteItemsd¶

class monai.transforms.DeleteItemsd(keys)[source]

Delete specified items from data dictionary to release memory. It will remove the key-values and copy the others to construct a new dictionary.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Returns

An updated dictionary version of data by applying the transform.

#### SelectItemsd¶

class monai.transforms.SelectItemsd(keys)[source]

Select only specified items from data dictionary to release memory. It will copy the selected key-values and construct and new dictionary.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Returns

An updated dictionary version of data by applying the transform.

#### SqueezeDimd¶

class monai.transforms.SqueezeDimd(keys, dim=0)[source]

Dictionary-based wrapper of monai.transforms.SqueezeDim.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• dim (int) – dimension to be squeezed. Default: 0 (the first dimension)

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, Union[ndarray, Tensor]]

Returns

An updated dictionary version of data by applying the transform.

#### DataStatsd¶

class monai.transforms.DataStatsd(keys, prefix='Data', data_shape=True, value_range=True, data_value=False, additional_info=None, logger_handler=None)[source]

Dictionary-based wrapper of monai.transforms.DataStats.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• prefix (Union[Sequence[str], str]) – will be printed in format: “{prefix} statistics”. it also can be a sequence of string, each element corresponds to a key in keys.

• data_shape (Union[Sequence[bool], bool]) – whether to show the shape of input data. it also can be a sequence of bool, each element corresponds to a key in keys.

• value_range (Union[Sequence[bool], bool]) – whether to show the value range of input data. it also can be a sequence of bool, each element corresponds to a key in keys.

• data_value (Union[Sequence[bool], bool]) – whether to show the raw value of input data. it also can be a sequence of bool, each element corresponds to a key in keys. a typical example is to print some properties of Nifti image: affine, pixdim, etc.

• additional_info (Union[Sequence[Callable], Callable, None]) – user can define callable function to extract additional info from input data. it also can be a sequence of string, each element corresponds to a key in keys.

• logger_handler (Optional[Handler]) – add additional handler to output data: save to file, etc. add existing python logging handlers: https://docs.python.org/3/library/logging.handlers.html

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, Union[ndarray, Tensor]]

Returns

An updated dictionary version of data by applying the transform.

#### SimulateDelayd¶

class monai.transforms.SimulateDelayd(keys, delay_time=0.0)[source]

Dictionary-based wrapper of monai.transforms.SimulateDelay.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• delay_time (Union[Sequence[float], float]) – The minimum amount of time, in fractions of seconds, to accomplish this identity task. It also can be a sequence of string, each element corresponds to a key in keys.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, Union[ndarray, Tensor]]

Returns

An updated dictionary version of data by applying the transform.

#### CopyItemsd¶

class monai.transforms.CopyItemsd(keys, times, names)[source]

Copy specified items from data dictionary and save with different key names. It can copy several items together and copy several times.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• times (int) – expected copy times, for example, if keys is “img”, times is 3, it will add 3 copies of “img” data to the dictionary.

• names (Union[Collection[Hashable], Hashable]) – the names corresponding to the newly copied data, the length should match len(keys) x times. for example, if keys is [“img”, “seg”] and times is 2, names can be: [“img_1”, “seg_1”, “img_2”, “seg_2”].

Raises
• ValueError – When times is nonpositive.

• ValueError – When len(names) is not len(keys) * times. Incompatible values.

__call__(data)[source]
Raises

KeyError – When a key in self.names already exists in data.

#### ConcatItemsd¶

class monai.transforms.ConcatItemsd(keys, name, dim=0)[source]

Concatenate specified items from data dictionary together on the first dim to construct a big array. Expect all the items are numpy array or PyTorch Tensor.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be concatenated together. See also: monai.transforms.compose.MapTransform

• name (str) – the name corresponding to the key to store the concatenated data.

• dim (int) – on which dimension to concatenate the items, default is 0.

Raises

ValueError – When insufficient keys are given (len(self.keys) < 2).

__call__(data)[source]
Raises
• TypeError – When items in data differ in type.

• TypeError – When the item type is not in Union[numpy.ndarray, torch.Tensor].

Dictionary-based wrapper of monai.transforms.Lambda.

For example:

input_data={'image': np.zeros((10, 2, 2)), 'label': np.ones((10, 2, 2))}
lambd = Lambdad(keys='label', func=lambda x: x[:4, :, :])
print(lambd(input_data)['label'].shape)
(4, 2, 2)
Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• func (Union[Sequence[Callable], Callable]) – Lambda/function to be applied. It also can be a sequence of Callable, each element corresponds to a key in keys.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Returns

An updated dictionary version of data by applying the transform.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• select_labels (Union[Sequence[int], int]) – labels to generate mask from. for 1 channel label, the select_labels is the expected label values, like: [1, 2, 3]. for One-Hot format label, the select_labels is the expected channel indices.

• merge_channels (bool) – whether to use np.any() to merge the result on channel dim. if yes, will return a single channel mask with binary data.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, ndarray]

Returns

An updated dictionary version of data by applying the transform.

#### FgBgToIndicesd¶

class monai.transforms.FgBgToIndicesd(keys, fg_postfix='_fg_indices', bg_postfix='_bg_indices', image_key=None, image_threshold=0.0, output_shape=None)[source]

Dictionary-based wrapper of monai.transforms.FgBgToIndices.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• fg_postfix (str) – postfix to save the computed foreground indices in dict. for example, if computed on label and postfix = “_fg_indices”, the key will be label_fg_indices.

• bg_postfix (str) – postfix to save the computed background indices in dict. for example, if computed on label and postfix = “_bg_indices”, the key will be label_bg_indices.

• image_key (Optional[str]) – if image_key is not None, use label == 0 & image > image_threshold to determine the negative sample(background). so the output items will not map to all the voxels in the label.

• image_threshold (float) – if enabled image_key, use image > image_threshold to determine the valid image content area and select background only in this area.

• output_shape (Optional[Sequence[int]]) – expected shape of output indices. if not None, unravel indices to specified shape.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, ndarray]

Returns

An updated dictionary version of data by applying the transform.

#### ConvertToMultiChannelBasedOnBratsClassesd¶

class monai.transforms.ConvertToMultiChannelBasedOnBratsClassesd(keys)[source]

Dictionary-based wrapper of monai.transforms.ConvertToMultiChannelBasedOnBratsClasses. Convert labels to multi channels based on brats18 classes: label 1 is the necrotic and non-enhancing tumor core label 2 is the the peritumoral edema label 4 is the GD-enhancing tumor The possible classes are TC (Tumor core), WT (Whole tumor) and ET (Enhancing tumor).

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, ndarray]

Returns

An updated dictionary version of data by applying the transform.

class monai.transforms.AddExtremePointsChanneld(keys, label_key, background=0, pert=0.0, sigma=3.0, rescale_min=- 1.0, rescale_max=1.0)[source]

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• label_key (str) – key to label source to get the extreme points.

• background (int) – Class index of background label, defaults to 0.

• pert (float) – Random perturbation amount to add to the points, defaults to 0.0.

• sigma (Union[Sequence[float], float, Sequence[Tensor], Tensor]) – if a list of values, must match the count of spatial dimensions of input data, and apply every value in the list to 1 spatial dimension. if only 1 value provided, use it for all spatial dimensions.

• rescale_min (float) – minimum value of output data.

• rescale_max (float) – maximum value of output data.

__call__(data)[source]

Call self as a function.

randomize(label)[source]

Within this method, self.R should be used, instead of np.random, to introduce random factors.

all self.R calls happen here so that we have a better chance to identify errors of sync the random state.

This method can generate the random factors based on properties of the input data.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

None

#### TorchVisiond¶

class monai.transforms.TorchVisiond(keys, name, *args, **kwargs)[source]

Dictionary-based wrapper of monai.transforms.TorchVision. As most of the TorchVision transforms only work for PIL image and PyTorch Tensor, this transform expects input data to be dict of PyTorch Tensors, users can easily call ToTensord transform to convert Numpy to Tensor.

Parameters
• keys (Union[Collection[Hashable], Hashable]) – keys of the corresponding items to be transformed. See also: monai.transforms.compose.MapTransform

• name (str) – The transform name in TorchVision package.

• args – parameters for the TorchVision transform.

• kwargs – parameters for the TorchVision transform.

__call__(data)[source]

data often comes from an iteration over an iterable, such as torch.utils.data.Dataset.

To simplify the input validations, this method assumes:

• data is a Python dictionary

• data[key] is a Numpy ndarray, PyTorch Tensor or string, where key is an element of self.keys, the data shape can be:

1. string data without shape, LoadImaged transform expects file paths

2. most of the pre-processing transforms expect: (num_channels, spatial_dim_1[, spatial_dim_2, ...]), except that AddChanneld expects (spatial_dim_1[, spatial_dim_2, …]) and AsChannelFirstd expects (spatial_dim_1[, spatial_dim_2, …], num_channels)

3. most of the post-processing transforms expect (batch_size, num_channels, spatial_dim_1[, spatial_dim_2, ...])

• the channel dimension is not omitted even if number of channels is one

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, Tensor]

Returns

An updated dictionary version of data by applying the transform.

### How to use the adaptor function¶

The key to using ‘adaptor’ lies in understanding the function that want to adapt. The ‘inputs’ and ‘outputs’ parameters take either strings, lists/tuples of strings or a dictionary mapping strings, depending on call signature of the function being called.

The adaptor function is written to minimise the cognitive load on the caller. There should be a minimal number of cases where the caller has to set anything on the input parameter, and for functions that return a single value, it is only necessary to name the dictionary keyword to which that value is assigned.

#### Use of outputs¶

outputs can take either a string, a list/tuple of string or a dict of string to string, depending on what the transform being adapted returns:

• If the transform returns a single argument, then outputs can be supplied a string that indicates what key to assign the return value to in the dictionary

• If the transform returns a list/tuple of values, then outputs can be supplied a list/tuple of the same length. The strings in outputs map the return value at the corresponding position to a key in the dictionary

• If the transform returns a dictionary of values, then outputs must be supplied a dictionary that maps keys in the function’s return dictionary to the dictionary being passed between functions

Note, the caller is free to use a more complex way of specifying the outputs parameter than is required. The following are synonymous and will be treated identically:

# single argument

# multiple arguments

#### Use of inputs¶

inputs can usually be omitted when using adaptor. It is only required when a the function’s parameter names do not match the names in the dictionary that is used to chain transform calls.

class MyTransform1:
def __call__(self, image):
# do stuff to image
return image + 1

class MyTransform2:
def __call__(self, img_dict):
# do stuff to image
img_dict["image"] += 1
return img_dict

d = {"image": 1}
print(xform(d))

>>> {'image': 3}
class MyTransform3:
def __call__(self, img_dict):
# do stuff to image
img_dict["image"] -= 1
img_dict["segment"] = img_dict["image"]
return img_dict

class MyTransform4:
def __call__(self, img, seg):
# do stuff to image
img -= 1
seg -= 1
return img, seg

xform = Compose([MyTransform3(), adaptor(MyTransform4(), ["img", "seg"], {"image": "img", "segment": "seg"})])
d = {"image": 1}
print(xform(d))

>>> {'image': 0, 'segment': 0, 'img': -1, 'seg': -1}

Inputs:

• dictionary in: None | Name maps

• params in (match): None | Name list | Name maps

• params in (mismatch): Name maps

• params & **kwargs (match) : None | Name maps

• params & **kwargs (mismatch) : Name maps

Outputs:

• dictionary out: None | Name maps

• list/tuple out: list/tuple

• variable out: string

## Utilities¶

monai.transforms.utils.apply_transform(transform, data, map_items=True)[source]

Transform data with transform. If data is a list or tuple and map_data is True, each item of data will be transformed and this method returns a list of outcomes. otherwise transform will be applied once with data as the argument.

Parameters
• transform (Callable) – a callable to be used to transform data

• data – an object to be transformed.

• map_items (bool) – whether to apply transform to each item in data, if data is a list or tuple. Defaults to True.

Raises

Exception – When transform raises an exception.

monai.transforms.utils.copypaste_arrays(src_shape, dest_shape, srccenter, destcenter, dims)[source]

Calculate the slices to copy a sliced area of array in src_shape into array in dest_shape.

The area has dimensions dims (use 0 or None to copy everything in that dimension), the source area is centered at srccenter index in src and copied into area centered at destcenter in dest. The dimensions of the copied area will be clipped to fit within the source and destination arrays so a smaller area may be copied than expected. Return value is the tuples of slice objects indexing the copied area in src, and those indexing the copy area in dest.

Example

src_shape = (6,6)
src = np.random.randint(0,10,src_shape)
dest = np.zeros_like(src)
srcslices, destslices = copypaste_arrays(src_shape, dest.shape, (3, 2),(2, 1),(3, 4))
dest[destslices] = src[srcslices]
print(src)
print(dest)

>>> [[9 5 6 6 9 6]
[4 3 5 6 1 2]
[0 7 3 2 4 1]
[3 0 0 1 5 1]
[9 4 7 1 8