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. thread-unsafe transforms should inherit monai.transforms.ThreadUnsafe.

  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 cause some memory issue or IPC sync issue, especially in the multi-processing environment of PyTorch DataLoader.

See Also

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.

backend: List[monai.utils.enums.TransformBackends] = []

Transforms should add data types to this list if they are capable of performing a transform without modifying the input type. For example, [“torch.Tensor”, “np.ndarray”] means that no copies of the data are required if the input is either “torch.Tensor” or “np.ndarray”.

MapTransform

class monai.transforms.MapTransform(keys, allow_missing_keys=False)[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:
            # raise exception unless allow_missing_keys==True.
    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.

key_iterator(data, *extra_iterables)[source]

Iterate across keys and optionally extra iterables. If key is missing, exception is raised if allow_missing_keys==False (default). If allow_missing_keys==True, key is skipped.

Parameters
  • data (Dict[Hashable, Any]) – data that the transform will be applied to

  • extra_iterables (Optional[Iterable]) – anything else to be iterated through

Return type

Generator

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 provides the flexibility of component-specific determinism without affecting the global states. It is recommended to use this API with monai.data.DataLoader for deterministic behaviour of the preprocessing pipelines. This API is not thread-safe. Additionally, deepcopying instance of this class often causes insufficient randomness as the random states will be duplicated.

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.

RandomizableTransform

class monai.transforms.RandomizableTransform(prob=1.0, do_transform=True)[source]

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

from monai.transforms import RandomizableTransform

class RandShiftIntensity100(RandomizableTransform):
    def randomize(self):
        super().randomize(None)
        self._offset = self.R.uniform(low=0, high=100)

    def __call__(self, img):
        self.randomize()
        if not self._do_transform:
            return img
        return img + self._offset

transform = RandShiftIntensity()
transform.set_random_state(seed=0)
print(transform(10))
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.

Return type

None

Compose

class monai.transforms.Compose(transforms=None, map_items=True, unpack_items=False)[source]

Compose provides the ability to chain a series of callables together in a sequential manner. Each transform in the sequence must take a single argument and return a single value.

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 that 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 takes a data item dictionary as input, and returns a sequence of data items in the transform chain, all following transforms will be applied to each item of this list if map_items is True (the default). If map_items is False, the returned sequence is passed whole to the next callable in the chain.

For example:

A Compose([transformA, transformB, transformC], map_items=True)(data_dict) could achieve the following patch-based transformation on the data_dict input:

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

  2. transformB crops out image patches from the ‘img’ and ‘seg’ of data_dict, and return a list of three patch samples:

    {'img': 3x100x100 data, 'seg': 1x100x100 data, 'shape': (100, 100)}
                         applying transformB
                             ---------->
    [{'img': 3x20x20 data, 'seg': 1x20x20 data, 'shape': (20, 20)},
     {'img': 3x20x20 data, 'seg': 1x20x20 data, 'shape': (20, 20)},
     {'img': 3x20x20 data, 'seg': 1x20x20 data, 'shape': (20, 20)},]
    
  3. transformC then randomly rotates or flips ‘img’ and ‘seg’ of each dictionary item in the list returned by transformB.

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.

flatten()[source]

Return a Composition with a simple list of transforms, as opposed to any nested Compositions.

e.g., t1 = Compose([x, x, x, x, Compose([Compose([x, x]), x, x])]).flatten() will result in the equivalent of t1 = Compose([x, x, x, x, x, x, x, x]).

inverse(data)[source]

Inverse of __call__.

Raises

NotImplementedError – When the subclass does not override this method.

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.

InvertibleTransform

class monai.transforms.InvertibleTransform[source]

Classes for invertible transforms.

This class exists so that an invert method can be implemented. This allows, for example, images to be cropped, rotated, padded, etc., during training and inference, and after be returned to their original size before saving to file for comparison in an external viewer.

When the __call__ method is called, the transformation information for each key is stored. If the transforms were applied to keys “image” and “label”, there will be two extra keys in the dictionary: “image_transforms” and “label_transforms”. Each list contains a list of the transforms applied to that key. When the inverse method is called, the inverse is called on each key individually, which allows for different parameters being passed to each label (e.g., different interpolation for image and label).

When the inverse method is called, the inverse transforms are applied in a last- in-first-out order. As the inverse is applied, its entry is removed from the list detailing the applied transformations. That is to say that during the forward pass, the list of applied transforms grows, and then during the inverse it shrinks back down to an empty list.

The information in data[key_transform] will be compatible with the default collate since it only stores strings, numbers and arrays.

We currently check that the id() of the transform is the same in the forward and inverse directions. This is a useful check to ensure that the inverses are being processed in the correct order. However, this may cause issues if the id() of the object changes (such as multiprocessing on Windows). If you feel this issue affects you, please raise a GitHub issue.

Note to developers: When converting a transform to an invertible transform, you need to:

  1. Inherit from this class.

  2. In __call__, add a call to push_transform.

  3. Any extra information that might be needed for the inverse can be included with the dictionary extra_info. This dictionary should have the same keys regardless of whether do_transform was True or False and can only contain objects that are accepted in pytorch data loader’s collate function (e.g., None is not allowed).

  4. Implement an inverse method. Make sure that after performing the inverse, pop_transform is called.

check_transforms_match(transform)[source]

Check transforms are of same instance.

Return type

None

get_most_recent_transform(data, key)[source]

Get most recent transform.

Return type

dict

inverse(data)[source]

Inverse of __call__.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

dict

pop_transform(data, key)[source]

Remove most recent transform.

Return type

None

push_transform(data, key, extra_info=None, orig_size=None)[source]

Append to list of applied transforms for that key.

Return type

None

BatchInverseTransform

class monai.transforms.BatchInverseTransform(transform, loader, collate_fn=<function no_collation>, num_workers=0, detach=True)[source]

Perform inverse on a batch of data. This is useful if you have inferred a batch of images and want to invert them all.

Parameters
  • transform (InvertibleTransform) – a callable data transform on input data.

  • loader (DataLoader) – data loader used to run transforms and generate the batch of data.

  • collate_fn (Optional[Callable]) – how to collate data after inverse transformations. default won’t do any collation, so the output will be a list of size batch size.

  • num_workers (Optional[int]) – number of workers when run data loader for inverse transforms, default to 0 as only run 1 iteration and multi-processing may be even slower. if the transforms are really slow, set num_workers for multi-processing. if set to None, use the num_workers of the transform data loader.

  • detach (bool) – whether to detach the tensors. Scalars tensors will be detached into number types instead of torch tensors.

__init__(transform, loader, collate_fn=<function no_collation>, num_workers=0, detach=True)[source]
Parameters
  • transform (InvertibleTransform) – a callable data transform on input data.

  • loader (DataLoader) – data loader used to run transforms and generate the batch of data.

  • collate_fn (Optional[Callable]) – how to collate data after inverse transformations. default won’t do any collation, so the output will be a list of size batch size.

  • num_workers (Optional[int]) – number of workers when run data loader for inverse transforms, default to 0 as only run 1 iteration and multi-processing may be even slower. if the transforms are really slow, set num_workers for multi-processing. if set to None, use the num_workers of the transform data loader.

  • detach (bool) – whether to detach the tensors. Scalars tensors will be detached into number types instead of torch tensors.

Decollated

class monai.transforms.Decollated(keys=None, detach=True, allow_missing_keys=False)[source]

Decollate a batch of data, if input a dictionary, it can also support to only decollate specified keys. Note that unlike most MapTransforms, it will delete other keys not specified and if keys=None, will decollate all the data in the input. And it replicates the scalar values to every item of the decollated list.

Parameters
  • keys (Union[Collection[Hashable], Hashable, None]) – keys of the corresponding items to decollate, note that it will delete other keys not specified. if None, will decollate all the keys. see also: monai.transforms.compose.MapTransform.

  • detach (bool) – whether to detach the tensors. Scalars tensors will be detached into number types instead of torch tensors.

  • allow_missing_keys (bool) – don’t raise exception if key is missing.

Vanilla Transforms

Crop and Pad

Pad

class monai.transforms.Pad(to_pad, mode=NumpyPadMode.CONSTANT, **kwargs)[source]

Perform padding for a given an amount of padding in each dimension. If input is torch.Tensor, torch.nn.functional.pad will be used, otherwise, np.pad will be used.

Parameters
  • to_pad (List[Tuple[int, int]]) – the amount to be padded in each dimension [(low_H, high_H), (low_W, high_W), …].

  • mode (Union[NumpyPadMode, PytorchPadMode, str]) – available modes for numpy array:{"constant", "edge", "linear_ramp", "maximum", "mean", "median", "minimum", "reflect", "symmetric", "wrap", "empty"} available modes for PyTorch Tensor: {"constant", "reflect", "replicate", "circular"}. 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 https://pytorch.org/docs/stable/generated/torch.nn.functional.pad.html

  • kwargs – other arguments for the np.pad or torch.pad function. note that np.pad treats channel dimension as the first dimension.

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

img (Union[ndarray, Tensor]) – data to be transformed, assuming img is channel-first and padding doesn’t apply to the channel dim.

mode: available modes for numpy array:{"constant", "edge", "linear_ramp", "maximum",

"mean", "median", "minimum", "reflect", "symmetric", "wrap", "empty"} available modes for PyTorch Tensor: {"constant", "reflect", "replicate" or "circular"}. 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 https://pytorch.org/docs/stable/generated/torch.nn.functional.pad.html

Return type

Union[ndarray, Tensor]

SpatialPad

class monai.transforms.SpatialPad(spatial_size, method=Method.SYMMETRIC, mode=NumpyPadMode.CONSTANT, **kwargs)[source]

Performs padding to the data, symmetric for all sides or all on one side for each dimension.

If input is torch.Tensor and mode is constant, torch.nn.functional.pad will be used. Otherwise, np.pad will be used (input converted to np.ndarray if necessary).

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 a dimension of the input data size is bigger than the pad size, will not pad that dimension. If its components have non-positive values, the corresponding size of input image will be used (no padding). for example: if the spatial size of input data is [30, 30, 30] and spatial_size=[32, 25, -1], the spatial size of output data will be [32, 30, 30].

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

  • mode (Union[NumpyPadMode, PytorchPadMode, str]) – available modes for numpy array:{"constant", "edge", "linear_ramp", "maximum", "mean", "median", "minimum", "reflect", "symmetric", "wrap", "empty"} available modes for PyTorch Tensor: {"constant", "reflect", "replicate", "circular"}. 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 https://pytorch.org/docs/stable/generated/torch.nn.functional.pad.html

  • kwargs – other arguments for the np.pad or torch.pad function. note that np.pad treats channel dimension as the first dimension.

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

  • mode (Union[NumpyPadMode, PytorchPadMode, str, None]) – available modes for numpy array:{"constant", "edge", "linear_ramp", "maximum", "mean", "median", "minimum", "reflect", "symmetric", "wrap", "empty"} available modes for PyTorch Tensor: {"constant", "reflect", "replicate", "circular"}. 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 https://pytorch.org/docs/stable/generated/torch.nn.functional.pad.html

Return type

Union[ndarray, Tensor]

BorderPad

class monai.transforms.BorderPad(spatial_border, mode=NumpyPadMode.CONSTANT, **kwargs)[source]

Pad the input data by adding specified borders to every dimension.

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

    specified size for every spatial border. Any -ve values will be set to 0. 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, PytorchPadMode, str]) – available modes for numpy array:{"constant", "edge", "linear_ramp", "maximum", "mean", "median", "minimum", "reflect", "symmetric", "wrap", "empty"} available modes for PyTorch Tensor: {"constant", "reflect", "replicate", "circular"}. 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 https://pytorch.org/docs/stable/generated/torch.nn.functional.pad.html

  • kwargs – other arguments for the np.pad or torch.pad function. note that np.pad treats channel dimension as the first dimension.

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

  • mode (Union[NumpyPadMode, PytorchPadMode, str, None]) – available modes for numpy array:{"constant", "edge", "linear_ramp", "maximum", "mean", "median", "minimum", "reflect", "symmetric", "wrap", "empty"} available modes for PyTorch Tensor: {"constant", "reflect", "replicate", "circular"}. 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 https://pytorch.org/docs/stable/generated/torch.nn.functional.pad.html

Raises
  • ValueError – When self.spatial_border does not contain ints.

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

Return type

Union[ndarray, Tensor]

DivisiblePad

class monai.transforms.DivisiblePad(k, mode=NumpyPadMode.CONSTANT, method=Method.SYMMETRIC, **kwargs)[source]

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, PytorchPadMode, str]) – available modes for numpy array:{"constant", "edge", "linear_ramp", "maximum", "mean", "median", "minimum", "reflect", "symmetric", "wrap", "empty"} available modes for PyTorch Tensor: {"constant", "reflect", "replicate", "circular"}. 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 https://pytorch.org/docs/stable/generated/torch.nn.functional.pad.html

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

  • kwargs – other arguments for the np.pad or torch.pad function. note that np.pad treats channel dimension as the first dimension.

See also monai.transforms.SpatialPad

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

  • mode (Union[NumpyPadMode, PytorchPadMode, str, None]) – available modes for numpy array:{"constant", "edge", "linear_ramp", "maximum", "mean", "median", "minimum", "reflect", "symmetric", "wrap", "empty"} available modes for PyTorch Tensor: {"constant", "reflect", "replicate", "circular"}. 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 https://pytorch.org/docs/stable/generated/torch.nn.functional.pad.html

Return type

Union[ndarray, Tensor]

__init__(k, mode=NumpyPadMode.CONSTANT, method=Method.SYMMETRIC, **kwargs)[source]
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, PytorchPadMode, str]) – available modes for numpy array:{"constant", "edge", "linear_ramp", "maximum", "mean", "median", "minimum", "reflect", "symmetric", "wrap", "empty"} available modes for PyTorch Tensor: {"constant", "reflect", "replicate", "circular"}. 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 https://pytorch.org/docs/stable/generated/torch.nn.functional.pad.html

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

  • kwargs – other arguments for the np.pad or torch.pad function. note that np.pad treats channel dimension as the first dimension.

See also monai.transforms.SpatialPad

SpatialCrop

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

General purpose cropper to produce sub-volume region of interest (ROI). If a dimension of the expected ROI size is bigger than the input image size, will not crop that dimension. So the cropped result may be smaller than the expected ROI, and the cropped results of several images may not have exactly the same shape. It can support to crop ND spatial (channel-first) data.

The cropped region can be parameterised in various ways:
  • a list of slices for each spatial dimension (allows for use of -ve indexing and None)

  • a spatial center and size

  • the start and end coordinates of the ROI

Parameters
  • roi_center (Union[Sequence[int], ndarray, Tensor, None]) – voxel coordinates for center of the crop ROI.

  • roi_size (Union[Sequence[int], ndarray, Tensor, None]) – size of the crop ROI, if a dimension of ROI size is bigger than image size, will not crop that dimension of the image.

  • roi_start (Union[Sequence[int], ndarray, Tensor, None]) – voxel coordinates for start of the crop ROI.

  • roi_end (Union[Sequence[int], ndarray, Tensor, None]) – voxel coordinates for end of the crop ROI, if a coordinate is out of image, use the end coordinate of image.

  • roi_slices (Optional[Sequence[slice]]) – list of slices for each of the spatial dimensions.

__call__(img)[source]

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

Return type

Union[ndarray, Tensor]

__init__(roi_center=None, roi_size=None, roi_start=None, roi_end=None, roi_slices=None)[source]
Parameters
  • roi_center (Union[Sequence[int], ndarray, Tensor, None]) – voxel coordinates for center of the crop ROI.

  • roi_size (Union[Sequence[int], ndarray, Tensor, None]) – size of the crop ROI, if a dimension of ROI size is bigger than image size, will not crop that dimension of the image.

  • roi_start (Union[Sequence[int], ndarray, Tensor, None]) – voxel coordinates for start of the crop ROI.

  • roi_end (Union[Sequence[int], ndarray, Tensor, None]) – voxel coordinates for end of the crop ROI, if a coordinate is out of image, use the end coordinate of image.

  • roi_slices (Optional[Sequence[slice]]) – list of slices for each of the spatial dimensions.

CenterSpatialCrop

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

Crop at the center of image with specified ROI size. If a dimension of the expected ROI size is bigger than the input image size, will not crop that dimension. So the cropped result may be smaller than the expected ROI, and the cropped results of several images may not have exactly the same shape.

Parameters

roi_size (Union[Sequence[int], int]) – the spatial size of the crop region e.g. [224,224,128] if a dimension of ROI size is bigger than image size, will not crop that dimension of the image. If its components have non-positive values, the corresponding size of input image will be used. for example: if the spatial size of input data is [40, 40, 40] and roi_size=[32, 64, -1], the spatial size of output data will be [32, 40, 40].

__call__(img)[source]

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

RandSpatialCrop

class monai.transforms.RandSpatialCrop(roi_size, max_roi_size=None, 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 and maximum size to limit the randomly generated ROI.

Note: even random_size=False, if a dimension of the expected ROI size is bigger than the input image size, will not crop that dimension. So the cropped result may be smaller than the expected ROI, and the cropped results of several images may not have exactly the same shape.

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 a dimension of ROI size is bigger than image size, will not crop that dimension of the image. If its components have non-positive values, the corresponding size of input image will be used. for example: if the spatial size of input data is [40, 40, 40] and roi_size=[32, 64, -1], the spatial size of output data will be [32, 40, 40].

  • max_roi_size (Union[Sequence[int], int, None]) – if random_size is True and roi_size specifies the min crop region size, max_roi_size can specify the max crop region size. if None, defaults to the input image size. 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. if True, the actual size is sampled from randint(roi_size, max_roi_size + 1).

__call__(img)[source]

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

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, max_roi_size=None, 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.

Note: even random_size=False, if a dimension of the expected ROI size is bigger than the input image size, will not crop that dimension. So the cropped result may be smaller than the expected ROI, and the cropped results of several images may not have exactly the same shape.

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 a dimension of ROI size is bigger than image size, will not crop that dimension of the image. If its components have non-positive values, the corresponding size of input image will be used. for example: if the spatial size of input data is [40, 40, 40] and roi_size=[32, 64, -1], the spatial size of output data will be [32, 40, 40].

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

  • max_roi_size (Union[Sequence[int], int, None]) – if random_size is True and roi_size specifies the min crop region size, max_roi_size can specify the max crop region size. if None, defaults to the input image size. 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).

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 is_positive>, channel_indices=None, margin=0, return_coords=False, k_divisible=1, mode=NumpyPadMode.CONSTANT, **np_kwargs)[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


def threshold_at_one(x):
    # threshold at 1
    return x > 1


cropper = CropForeground(select_fn=threshold_at_one, 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.

  • k_divisible (Union[Sequence[int], int]) – make each spatial dimension to be divisible by k, default to 1. if k_divisible is an int, the same k be applied to all the input spatial dimensions.

  • mode (Union[NumpyPadMode, str]) – padding mode {"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

  • np_kwargs – other args for np.pad API, note that np.pad treats channel dimension as the first dimension. more details: https://numpy.org/doc/1.18/reference/generated/numpy.pad.html

__call__(img, mode=None)[source]

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

__init__(select_fn=<function is_positive>, channel_indices=None, margin=0, return_coords=False, k_divisible=1, mode=NumpyPadMode.CONSTANT, **np_kwargs)[source]
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.

  • k_divisible (Union[Sequence[int], int]) – make each spatial dimension to be divisible by k, default to 1. if k_divisible is an int, the same k be applied to all the input spatial dimensions.

  • mode (Union[NumpyPadMode, str]) – padding mode {"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

  • np_kwargs – other args for np.pad API, note that np.pad treats channel dimension as the first dimension. more details: https://numpy.org/doc/1.18/reference/generated/numpy.pad.html

compute_bounding_box(img)[source]

Compute the start points and end points of bounding box to crop. And adjust bounding box coords to be divisible by k.

crop_pad(img, box_start, box_end, mode=None)[source]

Crop and pad based on the bounding box.

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]]]

If a dimension of the expected spatial size is bigger than the input image size, will not crop that dimension. So the cropped result may be smaller than expected size, and the cropped results of several images may not have exactly same shape.

Parameters
  • spatial_size (Union[Sequence[int], int]) – the spatial size of the crop region e.g. [224, 224, 128]. if a dimension of ROI size is bigger than image size, will not crop that dimension of the image. if its components have non-positive values, the corresponding size of label will be used. for example: if the spatial size of input data is [40, 40, 40] and spatial_size=[32, 64, -1], the spatial size of output data will be [32, 40, 40].

  • 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

RandCropByLabelClasses

class monai.transforms.RandCropByLabelClasses(spatial_size, ratios=None, label=None, num_classes=None, num_samples=1, image=None, image_threshold=0.0, indices=None)[source]

Crop random fixed sized regions with the center being a class based on the specified ratios of every class. The label data can be One-Hot format array or Argmax data. 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 ratios=[1, 2, 3, 1]:

image = np.array([
    [[0.0, 0.3, 0.4, 0.2, 0.0],
    [0.0, 0.1, 0.2, 0.1, 0.4],
    [0.0, 0.3, 0.5, 0.2, 0.0],
    [0.1, 0.2, 0.1, 0.1, 0.0],
    [0.0, 0.1, 0.2, 0.1, 0.0]]
])
label = np.array([
    [[0, 0, 0, 0, 0],
    [0, 1, 2, 1, 0],
    [0, 1, 3, 0, 0],
    [0, 0, 0, 0, 0],
    [0, 0, 0, 0, 0]]
])
cropper = RandCropByLabelClasses(
    spatial_size=[3, 3],
    ratios=[1, 2, 3, 1],
    num_classes=4,
    num_samples=2,
)
label_samples = cropper(img=label, label=label, image=image)

The 2 randomly cropped samples of `label` can be:
[[0, 1, 2],     [[0, 0, 0],
 [0, 1, 3],      [1, 2, 1],
 [0, 0, 0]]      [1, 3, 0]]

If a dimension of the expected spatial size is bigger than the input image size, will not crop that dimension. So the cropped result may be smaller than expected size, and the cropped results of several images may not have exactly same shape.

Parameters
  • spatial_size (Union[Sequence[int], int]) – the spatial size of the crop region e.g. [224, 224, 128]. if a dimension of ROI size is bigger than image size, will not crop that dimension of the image. if its components have non-positive values, the corresponding size of label will be used. for example: if the spatial size of input data is [40, 40, 40] and spatial_size=[32, 64, -1], the spatial size of output data will be [32, 40, 40].

  • ratios (Optional[List[Union[float, int]]]) – specified ratios of every class in the label to generate crop centers, including background class. if None, every class will have the same ratio to generate crop centers.

  • label (Optional[ndarray]) – the label image that is used for finding every classes, if None, must set at self.__call__.

  • num_classes (Optional[int]) – number of classes for argmax label, not necessary for One-Hot label.

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

  • image (Optional[ndarray]) – if image is not None, only return the indices of every class that are within the valid region of the image (image > image_threshold).

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

  • indices (Optional[List[ndarray]]) – if provided pre-computed indices of every class, will ignore above image and image_threshold, and randomly select crop centers based on them, expect to be 1 dim array of spatial indices after flattening. a typical usage is to call ClassesToIndices transform first and cache the results for better performance.

__call__(img, label=None, image=None, indices=None)[source]
Parameters
  • img (ndarray) – input data to crop samples from based on the ratios of every class, assumes img is a channel-first array.

  • label (Optional[ndarray]) – the label image that is used for finding indices of every class, 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 image > image_threshold to select the centers only in valid region. if None, use self.image.

  • indices (Optional[List[ndarray]]) – list of indices for every class in the image, used to randomly select crop centers.

Return type

List[ndarray]

randomize(label, 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

ResizeWithPadOrCrop

class monai.transforms.ResizeWithPadOrCrop(spatial_size, mode=NumpyPadMode.CONSTANT, method=Method.SYMMETRIC, **np_kwargs)[source]

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

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

  • np_kwargs – other args for np.pad API, note that np.pad treats channel dimension as the first dimension. more details: 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 is_positive>)[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]

See also: monai.transforms.utils.generate_spatial_bounding_box.

Return type

ndarray

RandScaleCrop

class monai.transforms.RandScaleCrop(roi_scale, max_roi_scale=None, random_center=True, random_size=True)[source]

Subclass of 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 and maximum scale of image size to limit the randomly generated ROI.

Parameters
  • roi_scale (Union[Sequence[float], float]) – if random_size is True, it specifies the minimum crop size: roi_scale * image spatial size. if random_size is False, it specifies the expected scale of image size to crop. e.g. [0.3, 0.4, 0.5]. If its components have non-positive values, will use 1.0 instead, which means the input image size.

  • max_roi_scale (Union[Sequence[float], float, None]) – if random_size is True and roi_scale specifies the min crop region size, max_roi_scale can specify the max crop region size: max_roi_scale * image spatial size. if None, defaults to the input image size. if its components have non-positive values, will use 1.0 instead, which means the input image size.

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

  • random_size (bool) – crop with random size or specified size ROI by roi_scale * image spatial size. if True, the actual size is sampled from randint(roi_scale * image spatial size, max_roi_scale * image spatial size + 1).

__call__(img)[source]

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

CenterScaleCrop

class monai.transforms.CenterScaleCrop(roi_scale)[source]

Crop at the center of image with specified scale of ROI size.

Parameters

roi_scale (Union[Sequence[float], float]) – specifies the expected scale of image size to crop. e.g. [0.3, 0.4, 0.5] or a number for all dims. If its components have non-positive values, will use 1.0 instead, which means the input image size.

__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.

Intensity

RandGaussianNoise

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

Add Gaussian noise to image.

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[ndarray, Tensor]

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.

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, offset=None)[source]

Apply the transform to img.

Return type

Union[ndarray, Tensor]

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, factor=None)[source]

Apply the transform to img.

Parameters
  • img (Union[ndarray, Tensor]) – input image to shift intensity.

  • factor (Optional[float]) – a factor to multiply the random offset, then shift. can be some image specific value at runtime, like: max(img), etc.

Return type

Union[ndarray, Tensor]

__init__(offsets, prob=0.1)[source]
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.

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.

Return type

None

StdShiftIntensity

class monai.transforms.StdShiftIntensity(factor, nonzero=False, channel_wise=False, dtype=<class 'numpy.float32'>)[source]

Shift intensity for the image with a factor and the standard deviation of the image by: v = v + factor * std(v). This transform can focus on only non-zero values or the entire image, and can also calculate the std on each channel separately.

Parameters
  • factor (float) – factor shift by v = v + factor * std(v).

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

  • channel_wise (bool) – if True, calculate on each channel separately. Please ensure that the first dimension represents the channel of the image if True.

  • dtype (Union[dtype, type, None]) – output data type, defaults to float32.

__call__(img)[source]

Apply the transform to img.

Return type

Union[ndarray, Tensor]

RandStdShiftIntensity

class monai.transforms.RandStdShiftIntensity(factors, prob=0.1, nonzero=False, channel_wise=False, dtype=<class 'numpy.float32'>)[source]

Shift intensity for the image with a factor and the standard deviation of the image by: v = v + factor * std(v) where the factor is randomly picked.

Parameters
  • factors (Union[Tuple[float, float], float]) – if tuple, the randomly picked range is (min(factors), max(factors)). If single number, the range is (-factors, factors).

  • prob (float) – probability of std shift.

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

  • channel_wise (bool) – if True, calculate on each channel separately.

  • dtype (Union[dtype, type, None]) – output data type, defaults to float32.

__call__(img)[source]

Apply the transform to img.

Return type

Union[ndarray, Tensor]

__init__(factors, prob=0.1, nonzero=False, channel_wise=False, dtype=<class 'numpy.float32'>)[source]
Parameters
  • factors (Union[Tuple[float, float], float]) – if tuple, the randomly picked range is (min(factors), max(factors)). If single number, the range is (-factors, factors).

  • prob (float) – probability of std shift.

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

  • channel_wise (bool) – if True, calculate on each channel separately.

  • dtype (Union[dtype, type, None]) – output data type, defaults to float32.

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.

Return type

None

RandBiasField

class monai.transforms.RandBiasField(degree=3, coeff_range=(0.0, 0.1), dtype=<class 'numpy.float32'>, prob=1.0)[source]

Random bias field augmentation for MR images. The bias field is considered as a linear combination of smoothly varying basis (polynomial) functions, as described in Automated Model-Based Tissue Classification of MR Images of the Brain. This implementation adapted from NiftyNet. Referred to Longitudinal segmentation of age-related white matter hyperintensities.

Parameters
  • degree (int) – degree of freedom of the polynomials. The value should be no less than 1. Defaults to 3.

  • coeff_range (Tuple[float, float]) – range of the random coefficients. Defaults to (0.0, 0.1).

  • dtype (Union[dtype, type, None]) – output data type, defaults to float32.

  • prob (float) – probability to do random bias field.

__call__(img)[source]

Apply the transform to img.

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.

Return type

None

ScaleIntensity

class monai.transforms.ScaleIntensity(minv=0.0, maxv=1.0, factor=None, dtype=<class 'numpy.float32'>)[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). In order to use this parameter, please set minv and maxv into None.

  • dtype (Union[dtype, type, None]) – output data type, defaults to float32.

__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

Union[ndarray, Tensor]

__init__(minv=0.0, maxv=1.0, factor=None, dtype=<class 'numpy.float32'>)[source]
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). In order to use this parameter, please set minv and maxv into None.

  • dtype (Union[dtype, type, None]) – output data type, defaults to float32.

RandScaleIntensity

class monai.transforms.RandScaleIntensity(factors, prob=0.1, dtype=<class 'numpy.float32'>)[source]

Randomly scale the intensity of input image by v = v * (1 + factor) where the factor is randomly picked.

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.

  • dtype (Union[dtype, type, None]) – output data type, defaults to float32.

__call__(img)[source]

Apply the transform to img.

Return type

Union[ndarray, Tensor]

__init__(factors, prob=0.1, dtype=<class 'numpy.float32'>)[source]
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.

  • dtype (Union[dtype, type, None]) – output data type, defaults to float32.

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.

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 (Union[Sequence, ndarray, Tensor, None]) – the amount to subtract by (usually the mean).

  • divisor (Union[Sequence, ndarray, Tensor, None]) – 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 (Union[dtype, type, None]) – output data type, defaults 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

Union[ndarray, Tensor]

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

Union[ndarray, Tensor]

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

Union[ndarray, Tensor]

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

Union[ndarray, Tensor]

AdjustContrast

class monai.transforms.AdjustContrast(gamma)[source]

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.

RandAdjustContrast

class monai.transforms.RandAdjustContrast(prob=0.1, gamma=(0.5, 4.5))[source]

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.

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.

Return type

None

MaskIntensity

class monai.transforms.MaskIntensity(mask_data=None, select_fn=<function is_positive>)[source]

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 the selected values in the mask data will keep the original value, others will be set to 0.

Parameters
  • mask_data (Optional[ndarray]) – if mask_data is single channel, apply to every channel of input image. if multiple channels, the number of channels must match the input data. the intensity values of input image corresponding to the selected values in the mask data will keep the original value, others will be set to 0. if None, must specify the mask_data at runtime.

  • select_fn (Callable) – function to select valid values of the mask_data, default is to select values > 0.

__call__(img, mask_data=None)[source]
Parameters

mask_data (Optional[ndarray]) – if mask data is single channel, apply to every 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 both mask_data and self.mask_data are None.

  • - 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 (Union[ndarray, Tensor]) – array containing input data. Must be real and in shape [channels, spatial1, spatial2, …].

Return type

Tensor

Returns

array 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.

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]

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.

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.

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.

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]

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.

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.

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]

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

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.

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, …].

Returns

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

GibbsNoise

class monai.transforms.GibbsNoise(alpha=0.5, as_tensor_output=True)[source]

The transform applies Gibbs noise to 2D/3D MRI images. Gibbs artifacts are one of the common type of type artifacts appearing in MRI scans.

The transform is applied to all the channels in the data.

For general information on Gibbs artifacts, please refer to:

An Image-based Approach to Understanding the Physics of MR Artifacts.

The AAPM/RSNA Physics Tutorial for Residents

Parameters
  • alpha (float) – Parametrizes the intensity of the Gibbs noise filter applied. Takes values in the interval [0,1] with alpha = 0 acting as the identity mapping.

  • as_tensor_output (bool) – if true return torch.Tensor, else return np.array. Default: True.

__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

Union[Tensor, ndarray]

RandGibbsNoise

class monai.transforms.RandGibbsNoise(prob=0.1, alpha=(0.0, 1.0), as_tensor_output=True)[source]

Naturalistic image augmentation via Gibbs artifacts. The transform randomly applies Gibbs noise to 2D/3D MRI images. Gibbs artifacts are one of the common type of type artifacts appearing in MRI scans.

The transform is applied to all the channels in the data.

For general information on Gibbs artifacts, please refer to: https://pubs.rsna.org/doi/full/10.1148/rg.313105115 https://pubs.rsna.org/doi/full/10.1148/radiographics.22.4.g02jl14949

Parameters
  • prob (float) – probability of applying the transform.

  • alpha (float, Sequence(float)) – Parametrizes the intensity of the Gibbs noise filter applied. Takes values in the interval [0,1] with alpha = 0 acting as the identity mapping. If a length-2 list is given as [a,b] then the value of alpha will be sampled uniformly from the interval [a,b]. 0 <= a <= b <= 1.

  • as_tensor_output (bool) – if true return torch.Tensor, else return np.array. default: True.

__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

Union[Tensor, ndarray]

KSpaceSpikeNoise

class monai.transforms.KSpaceSpikeNoise(loc, k_intensity=None, as_tensor_output=True)[source]

Apply localized spikes in k-space at the given locations and intensities. Spike (Herringbone) artifact is a type of data acquisition artifact which may occur during MRI scans.

For general information on spike artifacts, please refer to:

AAPM/RSNA physics tutorial for residents: fundamental physics of MR imaging.

Body MRI artifacts in clinical practice: A physicist’s and radiologist’s perspective.

Parameters
  • loc (Union[Tuple, Sequence[Tuple]]) – spatial location for the spikes. For images with 3D spatial dimensions, the user can provide (C, X, Y, Z) to fix which channel C is affected, or (X, Y, Z) to place the same spike in all channels. For 2D cases, the user can provide (C, X, Y) or (X, Y).

  • k_intensity (Union[Sequence[float], float, None]) – value for the log-intensity of the k-space version of the image. If one location is passed to loc or the channel is not specified, then this argument should receive a float. If loc is given a sequence of locations, then this argument should receive a sequence of intensities. This value should be tested as it is data-dependent. The default values are the 2.5 the mean of the log-intensity for each channel.

  • as_tensor_output (bool) – if True return torch.Tensor, else return np.array. Default: True.

Example

When working with 4D data, KSpaceSpikeNoise(loc = ((3,60,64,32), (64,60,32)), k_intensity = (13,14)) will place a spike at [3, 60, 64, 32] with log-intensity = 13, and one spike per channel located respectively at [: , 64, 60, 32] with log-intensity = 14.

__call__(img)[source]
Parameters

img (Union[ndarray, Tensor]) – image with dimensions (C, H, W) or (C, H, W, D)

Return type

Union[Tensor, ndarray]

RandKSpaceSpikeNoise

class monai.transforms.RandKSpaceSpikeNoise(prob=0.1, intensity_range=None, channel_wise=True, as_tensor_output=True)[source]

Naturalistic data augmentation via spike artifacts. The transform applies localized spikes in k-space, and it is the random version of monai.transforms.KSpaceSpikeNoise.

Spike (Herringbone) artifact is a type of data acquisition artifact which may occur during MRI scans. For general information on spike artifacts, please refer to:

AAPM/RSNA physics tutorial for residents: fundamental physics of MR imaging.

Body MRI artifacts in clinical practice: A physicist’s and radiologist’s perspective.

Parameters
  • prob (float) – probability of applying the transform, either on all channels at once, or channel-wise if channel_wise = True.

  • intensity_range (Optional[Sequence[Union[Sequence[float], float]]]) – pass a tuple (a, b) to sample the log-intensity from the interval (a, b) uniformly for all channels. Or pass sequence of intevals ((a0, b0), (a1, b1), …) to sample for each respective channel. In the second case, the number of 2-tuples must match the number of channels. Default ranges is (0.95x, 1.10x) where x is the mean log-intensity for each channel.

  • channel_wise – treat each channel independently. True by default.

  • as_tensor_output (bool) – if True return torch.Tensor, else return np.array. default: True.

Example

To apply k-space spikes randomly with probability 0.5, and log-intensity sampled from the interval [11, 12] for each channel independently, one uses RandKSpaceSpikeNoise(prob=0.5, intensity_range=(11, 12), channel_wise=True)

__call__(img)[source]

Apply transform to img. Assumes data is in channel-first form.

Parameters

img (Union[ndarray, Tensor]) – image with dimensions (C, H, W) or (C, H, W, D)

Return type

Union[Tensor, ndarray]

RandCoarseTransform

class monai.transforms.RandCoarseTransform(holes, spatial_size, max_holes=None, max_spatial_size=None, prob=0.1)[source]

Randomly select coarse regions in the image, then execute transform operations for the regions. It’s the base class of all kinds of region transforms. Refer to papers: https://arxiv.org/abs/1708.04552

Parameters
  • holes (int) – number of regions to dropout, if max_holes is not None, use this arg as the minimum number to randomly select the expected number of regions.

  • spatial_size (Union[Sequence[int], int]) – spatial size of the regions to dropout, if max_spatial_size is not None, use this arg as the minimum spatial size to randomly select size for every region. if some components of the spatial_size are non-positive values, the transform will use the corresponding components of input img size. For example, spatial_size=(32, -1) will be adapted to (32, 64) if the second spatial dimension size of img is 64.

  • max_holes (Optional[int]) – if not None, define the maximum number to randomly select the expected number of regions.

  • max_spatial_size (Union[Sequence[int], int, None]) – if not None, define the maximum spatial size to randomly select size for every region. if some components of the max_spatial_size are non-positive values, the transform will use the corresponding components of input img size. For example, max_spatial_size=(32, -1) will be adapted to (32, 64) if the second spatial dimension size of img is 64.

  • prob (float) – probability of applying the transform.

__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.

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.

Return type

None

RandCoarseDropout

class monai.transforms.RandCoarseDropout(holes, spatial_size, dropout_holes=True, fill_value=None, max_holes=None, max_spatial_size=None, prob=0.1)[source]

Randomly coarse dropout regions in the image, then fill in the rectangular regions with specified value. Or keep the rectangular regions and fill in the other areas with specified value. Refer to papers: https://arxiv.org/abs/1708.04552, https://arxiv.org/pdf/1604.07379 And other implementation: https://albumentations.ai/docs/api_reference/augmentations/transforms/ #albumentations.augmentations.transforms.CoarseDropout.

Parameters
  • holes (int) – number of regions to dropout, if max_holes is not None, use this arg as the minimum number to randomly select the expected number of regions.

  • spatial_size (Union[Sequence[int], int]) – spatial size of the regions to dropout, if max_spatial_size is not None, use this arg as the minimum spatial size to randomly select size for every region. if some components of the spatial_size are non-positive values, the transform will use the corresponding components of input img size. For example, spatial_size=(32, -1) will be adapted to (32, 64) if the second spatial dimension size of img is 64.

  • dropout_holes (bool) – if True, dropout the regions of holes and fill value, if False, keep the holes and dropout the outside and fill value. default to True.

  • fill_value (Union[Tuple[float, float], float, None]) – target value to fill the dropout regions, if providing a number, will use it as constant value to fill all the regions. if providing a tuple for the min and max, will randomly select value for every pixel / voxel from the range [min, max). if None, will compute the min and max value of input image then randomly select value to fill, default to None.

  • max_holes (Optional[int]) – if not None, define the maximum number to randomly select the expected number of regions.

  • max_spatial_size (Union[Sequence[int], int, None]) – if not None, define the maximum spatial size to randomly select size for every region. if some components of the max_spatial_size are non-positive values, the transform will use the corresponding components of input img size. For example, max_spatial_size=(32, -1) will be adapted to (32, 64) if the second spatial dimension size of img is 64.

  • prob (float) – probability of applying the transform.

RandCoarseShuffle

class monai.transforms.RandCoarseShuffle(holes, spatial_size, max_holes=None, max_spatial_size=None, prob=0.1)[source]

Randomly select regions in the image, then shuffle the pixels within every region. It shuffles every channel separately. Refer to paper: Kang, Guoliang, et al. “Patchshuffle regularization.” arXiv preprint arXiv:1707.07103 (2017). https://arxiv.org/abs/1707.07103

Parameters
  • holes (int) – number of regions to dropout, if max_holes is not None, use this arg as the minimum number to randomly select the expected number of regions.

  • spatial_size (Union[Sequence[int], int]) – spatial size of the regions to dropout, if max_spatial_size is not None, use this arg as the minimum spatial size to randomly select size for every region. if some components of the spatial_size are non-positive values, the transform will use the corresponding components of input img size. For example, spatial_size=(32, -1) will be adapted to (32, 64) if the second spatial dimension size of img is 64.

  • max_holes (Optional[int]) – if not None, define the maximum number to randomly select the expected number of regions.

  • max_spatial_size (Union[Sequence[int], int, None]) – if not None, define the maximum spatial size to randomly select size for every region. if some components of the max_spatial_size are non-positive values, the transform will use the corresponding components of input img size. For example, max_spatial_size=(32, -1) will be adapted to (32, 64) if the second spatial dimension size of img is 64.

  • prob (float) – probability of applying the transform.

HistogramNormalize

class monai.transforms.HistogramNormalize(num_bins=256, min=0, max=255, mask=None, dtype=<class 'numpy.float32'>)[source]

Apply the histogram normalization to input image. Refer to: https://github.com/facebookresearch/CovidPrognosis/blob/master/covidprognosis/data/transforms.py#L83.

Parameters
  • num_bins (int) – number of the bins to use in histogram, default to 256. for more details: https://numpy.org/doc/stable/reference/generated/numpy.histogram.html.

  • min (int) – the min value to normalize input image, default to 0.

  • max (int) – the max value to normalize input image, default to 255.

  • mask (Optional[ndarray]) – if provided, must be ndarray of bools or 0s and 1s, and same shape as image. only points at which mask==True are used for the equalization. can also provide the mask along with img at runtime.

  • dtype (Union[dtype, type, None]) – data type of the output, default to float32.

__call__(img, mask=None)[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

IO

LoadImage

class monai.transforms.LoadImage(reader=None, image_only=False, dtype=<class 'numpy.float32'>, *args, **kwargs)[source]

Load image file or files from provided path based on reader. If reader is not specified, this class automatically chooses readers based on the supported suffixes and in the following order:

  • User-specified reader at runtime when calling this loader.

  • User-specified reader in the constructor of LoadImage.

  • Readers from the last to the first in the registered list.

  • Current default readers: (nii, nii.gz -> NibabelReader), (png, jpg, bmp -> PILReader), (npz, npy -> NumpyReader), (others -> ITKReader).

Parameters
  • reader

    reader to load image file and meta data

    • if reader is None, a default set of SUPPORTED_READERS will be used.

    • if reader is a string, the corresponding item in SUPPORTED_READERS will be used, and a reader instance will be constructed with the *args and **kwargs parameters. the supported reader names are: “nibabelreader”, “pilreader”, “itkreader”, “numpyreader”.

    • if reader is a reader class/instance, it will be registered to this loader accordingly.

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

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

  • args – additional parameters for reader if providing a reader name.

  • kwargs – additional parameters for reader if providing a reader name.

Note

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

  • If reader is specified, the loader will attempt to use the specified readers and the default supported readers. This might introduce overheads when handling the exceptions of trying the incompatible loaders. In this case, it is therefore recommended to set the most appropriate reader as the last item of the reader parameter.

__call__(filename, reader=None)[source]

Load image file and meta data from the given filename(s). If reader is not specified, this class automatically chooses readers based on the reversed order of registered readers self.readers.

Parameters
  • filename (Union[Sequence[str], str, Path, Sequence[Path]]) – 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 to save, and will stack them together as multi-channels data. if provided directory path instead of file path, will treat it as DICOM images series and read.

  • reader (Optional[ImageReader]) – runtime reader to load image file and meta data.

__init__(reader=None, image_only=False, dtype=<class 'numpy.float32'>, *args, **kwargs)[source]
Parameters
  • reader

    reader to load image file and meta data

    • if reader is None, a default set of SUPPORTED_READERS will be used.

    • if reader is a string, the corresponding item in SUPPORTED_READERS will be used, and a reader instance will be constructed with the *args and **kwargs parameters. the supported reader names are: “nibabelreader”, “pilreader”, “itkreader”, “numpyreader”.

    • if reader is a reader class/instance, it will be registered to this loader accordingly.

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

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

  • args – additional parameters for reader if providing a reader name.

  • kwargs – additional parameters for reader if providing a reader name.

Note

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

  • If reader is specified, the loader will attempt to use the specified readers and the default supported readers. This might introduce overheads when handling the exceptions of trying the incompatible loaders. In this case, it is therefore recommended to set the most appropriate reader as the last item of the reader parameter.

register(reader)[source]

Register image reader to load image file and meta data.

Parameters

reader (ImageReader) – reader instance to be registered with this loader.

SaveImage

class monai.transforms.SaveImage(output_dir='./', output_postfix='trans', output_ext='.nii.gz', resample=True, mode='nearest', padding_mode=GridSamplePadMode.BORDER, scale=None, dtype=<class 'numpy.float64'>, output_dtype=<class 'numpy.float32'>, squeeze_end_dims=True, data_root_dir='', separate_folder=True, print_log=True)[source]

Save transformed data into files, support NIfTI and PNG formats. It can work for both numpy array and PyTorch Tensor in both preprocessing transform chain and postprocessing transform chain. The name of saved file will be {input_image_name}_{output_postfix}{output_ext}, where the input image name is extracted from the provided meta data dictionary. If no meta data provided, use index from 0 as the filename prefix. It can also save a list of PyTorch Tensor or numpy array without batch dim.

Note: image should be channel-first shape: [C,H,W,[D]].

Parameters
  • output_dir (Union[Path, str]) – output image directory.

  • output_postfix (str) – a string appended to all output file names, default to trans.

  • output_ext (str) – output file extension name, available extensions: .nii.gz, .nii, .png.

  • resample (bool) – whether to resample before saving the data array. if saving PNG format image, based on the spatial_shape from metadata. if saving NIfTI format image, based on the original_affine from metadata.

  • mode (Union[GridSampleMode, InterpolateMode, str]) –

    This option is used when resample = True. Defaults to "nearest".

  • padding_mode (Union[GridSamplePadMode, str]) –

    This option is used when resample = True. Defaults to "border".

  • scale (Optional[int]) – {255, 65535} postprocess data by clipping to [0, 1] and scaling [0, 255] (uint8) or [0, 65535] (uint16). Default is None to disable scaling. it’s used for PNG format only.

  • dtype (Union[dtype, type, None]) – data type during 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’s used for NIfTI format only.

  • output_dtype (Union[dtype, type, None]) – data type for saving data. Defaults to np.float32. it’s used for NIfTI format only.

  • squeeze_end_dims (bool) – if True, any trailing singleton dimensions will be removed (after the channel has been moved to the end). So if input is (C,H,W,D), this will be altered to (H,W,D,C), and then if C==1, it will be saved as (H,W,D). If D also ==1, it will be saved as (H,W). If false, image will always be saved as (H,W,D,C). it’s used for NIfTI format only.

  • data_root_dir (str) – if not empty, it specifies the beginning parts of the input file’s absolute path. it’s used to compute input_file_rel_path, the relative path to the file from data_root_dir to preserve folder structure when saving in case there are files in different folders with the same file names. for example: input_file_name: /foo/bar/test1/image.nii, output_postfix: seg output_ext: nii.gz output_dir: /output, data_root_dir: /foo/bar, output will be: /output/test1/image/image_seg.nii.gz

  • separate_folder (bool) – whether to save every file in a separate folder, for example: if input filename is image.nii, postfix is seg and folder_path is output, if True, save as: output/image/image_seg.nii, if False, save as output/image_seg.nii. default to True.

  • print_log (bool) – whether to print log about the saved file path, etc. default to True.

__call__(img, meta_data=None)[source]
Parameters
  • img (Union[Tensor, ndarray]) – target data content that save into file.

  • meta_data (Optional[Dict]) – key-value pairs of meta_data corresponding to the data.

NVIDIA Tool Extension (NVTX)

RangePush

class monai.transforms.RangePush(msg)[source]

Pushes a range onto a stack of nested range span. Stores zero-based depth of the range that is started.

Parameters

msg (str) – ASCII message to associate with range

RandRangePush

class monai.transforms.RandRangePush(msg)[source]

Pushes a range onto a stack of nested range span (RandomizableTransform). Stores zero-based depth of the range that is started.

Parameters

msg (str) – ASCII message to associate with range

RangePop

class monai.transforms.RangePop[source]

Pops a range off of a stack of nested range spans. Stores zero-based depth of the range that is ended.

RandRangePop

class monai.transforms.RandRangePop(prob=1.0, do_transform=True)[source]

Pops a range off of a stack of nested range spans (RandomizableTransform). Stores zero-based depth of the range that is ended.

Mark

class monai.transforms.Mark(msg)[source]

Mark an instantaneous event that occurred at some point.

Parameters

msg (str) – ASCII message to associate with the event.

RandMark

class monai.transforms.RandMark(msg)[source]

Mark an instantaneous event that occurred at some point. (RandomizableTransform)

Parameters

msg (str) – ASCII message to associate with the event.

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 = torch.tanh. 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, num_classes=None, threshold_values=False, logit_thresh=0.5, rounding=None, n_classes=None)[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.

  • round the value to the closest integer.

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.

  • num_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.

  • rounding (Optional[str]) – if not None, round the data according to the specified option, available options: [“torchrounding”].

__call__(img, argmax=None, to_onehot=None, num_classes=None, threshold_values=None, logit_thresh=None, rounding=None, n_classes=None)[source]
Parameters
  • img (Tensor) – the input tensor data to convert, if no channel dimension when converting to One-Hot, will automatically add it.

  • 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.

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

  • rounding (Optional[str]) – if not None, round the data according to the specified option, available options: [“torchrounding”].

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 channel-first PyTorch Tensor:
  1. With shape (1, spatial_dim1[, spatial_dim2, …]) and the values correspond to expected labels.

  2. With shape (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 (C, spatial_dim1[, spatial_dim2, …]).

Return type

Tensor

Returns

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

__init__(applied_labels, independent=True, connectivity=None)[source]
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.

LabelFilter

class monai.transforms.LabelFilter(applied_labels)[source]

This transform filters out labels and can be used as a processing step to view only certain labels.

The list of applied labels defines which labels will be kept.

Note

All labels which do not match the applied_labels are set to the background label (0).

For example:

Use LabelFilter with applied_labels=[1, 5, 9]:

[1, 2, 3]         [1, 0, 0]
[4, 5, 6]    =>   [0, 5 ,0]
[7, 8, 9]         [0, 0, 9]

Initialize the LabelFilter class with the labels to filter on.

Parameters

applied_labels (Union[Iterable[int], int]) – Label(s) to filter on.

__call__(img)[source]

Filter the image on the applied_labels.

Parameters

img (~NdarrayTensor) – Pytorch tensor or numpy array of any shape.

Raises

NotImplementedError – The provided image was not a Pytorch Tensor or numpy array.

Return type

~NdarrayTensor

Returns

Pytorch tensor or numpy array of the same shape as the input.

__init__(applied_labels)[source]

Initialize the LabelFilter class with the labels to filter on.

Parameters

applied_labels (Union[Iterable[int], int]) – Label(s) to filter on.

FillHoles

class monai.transforms.FillHoles(applied_labels=None, connectivity=None)[source]

This transform fills holes in the image and can be used to remove artifacts inside segments.

An enclosed hole is defined as a background pixel/voxel which is only enclosed by a single class. The definition of enclosed can be defined with the connectivity parameter:

1-connectivity     2-connectivity     diagonal connection close-up

     [ ]           [ ]  [ ]  [ ]             [ ]
      |               \  |  /                 |  <- hop 2
[ ]--[x]--[ ]      [ ]--[x]--[ ]        [x]--[ ]
      |               /  |  \             hop 1
     [ ]           [ ]  [ ]  [ ]

It is possible to define for which labels the hole filling should be applied. The input image is assumed to be a PyTorch Tensor or numpy array with shape [C, spatial_dim1[, spatial_dim2, …]]. If C = 1, then the values correspond to expected labels. If C > 1, then a one-hot-encoding is expected where the index of C matches the label indexing.

Note

The label 0 will be treated as background and the enclosed holes will be set to the neighboring class label.

The performance of this method heavily depends on the number of labels. It is a bit faster if the list of applied_labels is provided. Limiting the number of applied_labels results in a big decrease in processing time.

For example:

Use FillHoles with default parameters:

[1, 1, 1, 2, 2, 2, 3, 3]         [1, 1, 1, 2, 2, 2, 3, 3]
[1, 0, 1, 2, 0, 0, 3, 0]    =>   [1, 1 ,1, 2, 0, 0, 3, 0]
[1, 1, 1, 2, 2, 2, 3, 3]         [1, 1, 1, 2, 2, 2, 3, 3]

The hole in label 1 is fully enclosed and therefore filled with label 1. The background label near label 2 and 3 is not fully enclosed and therefore not filled.

Initialize the connectivity and limit the labels for which holes are filled.

Parameters
  • applied_labels (Union[Iterable[int], int, None]) – Labels for which to fill holes. Defaults to None, that is filling holes for all labels.

  • 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. Defaults to a full connectivity of input.ndim.

__call__(img)[source]

Fill the holes in the provided image.

Note

The value 0 is assumed as background label.

Parameters

img (~NdarrayTensor) – Pytorch Tensor or numpy array of shape [C, spatial_dim1[, spatial_dim2, …]].

Raises

NotImplementedError – The provided image was not a Pytorch Tensor or numpy array.

Return type

~NdarrayTensor

Returns

Pytorch Tensor or numpy array of shape [C, spatial_dim1[, spatial_dim2, …]].

__init__(applied_labels=None, connectivity=None)[source]

Initialize the connectivity and limit the labels for which holes are filled.

Parameters
  • applied_labels (Union[Iterable[int], int, None]) – Labels for which to fill holes. Defaults to None, that is filling holes for all labels.

  • 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. Defaults to a full connectivity of input.ndim.

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: [channels, height, width[, depth]]

Raises

ValueError – When image ndim is not one of [3, 4].

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: [C[, H, W, D]], Or a single PyTorch Tensor with shape: [E, 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, 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 2 dimensions, it will be added to E 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, 4, H, W, D]. and add different weights for different classes, so the weights shape can be: [3, 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

Prob NMS

class monai.transforms.ProbNMS(spatial_dims=2, sigma=0.0, prob_threshold=0.5, box_size=48)[source]

Performs probability based non-maximum suppression (NMS) on the probabilities map via iteratively selecting the coordinate with highest probability and then move it as well as its surrounding values. The remove range is determined by the parameter box_size. If multiple coordinates have the same highest probability, only one of them will be selected.

Parameters
  • spatial_dims (int) – number of spatial dimensions of the input probabilities map. Defaults to 2.

  • sigma (Union[Sequence[float], float, Sequence[Tensor], Tensor]) – the standard deviation for gaussian filter. It could be a single value, or spatial_dims number of values. Defaults to 0.0.

  • prob_threshold (float) – the probability threshold, the function will stop searching if the highest probability is no larger than the threshold. The value should be no less than 0.0. Defaults to 0.5.

  • box_size (Union[int, Sequence[int]]) – the box size (in pixel) to be removed around the the pixel with the maximum probability. It can be an integer that defines the size of a square or cube, or a list containing different values for each dimensions. Defaults to 48.

Returns

a list of selected lists, where inner lists contain probability and coordinates. For example, for 3D input, the inner lists are in the form of [probability, x, y, z].

Raises
  • ValueError – When prob_threshold is less than 0.0.

  • ValueError – When box_size is a list or tuple, and its length is not equal to spatial_dims.

  • ValueError – When box_size has a less than 1 value.

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: [C[, H, W, D]], Or a single PyTorch Tensor with shape: [E[, 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

class monai.transforms.Spacing(pixdim, diagonal=False, mode=GridSampleMode.BILINEAR, padding_mode=GridSamplePadMode.BORDER, align_corners=False, dtype=<class 'numpy.float64'>)[source]

Resample input image into the specified pixdim.

Parameters
  • pixdim (Union[Sequence[float], float]) – output voxel spacing. if providing a single number, will use it for the first dimension. items of the pixdim sequence map to the spatial dimensions of input image, if length of pixdim sequence is longer than image spatial dimensions, will ignore the longer part, if shorter, will pad with 1.0. if the components of the pixdim are non-positive values, the transform will use the corresponding components of the original pixdim, which is computed from the affine matrix of input image.

  • 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

  • padding_mode (Union[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

  • 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 (Union[dtype, type, 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.

__call__(data_array, affine=None, mode=None, padding_mode=None, align_corners=None, dtype=None, output_spatial_shape=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

  • padding_mode (Union[GridSamplePadMode, str, None]) – {"zeros", "border", "reflection"} Padding mode for outside grid values. Defaults to self.padding_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 (Union[dtype, type, None]) – 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.

  • output_spatial_shape (Optional[ndarray]) – specify the shape of the output data_array. This is typically useful for the inverse of Spacingd where sometimes we could not compute the exact shape due to the quantization error with the affine.

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 affine, current affine.

__init__(pixdim, diagonal=False, mode=GridSampleMode.BILINEAR, padding_mode=GridSamplePadMode.BORDER, align_corners=False, dtype=<class 'numpy.float64'>)[source]
Parameters
  • pixdim (Union[Sequence[float], float]) – output voxel spacing. if providing a single number, will use it for the first dimension. items of the pixdim sequence map to the spatial dimensions of input image, if length of pixdim sequence is longer than image spatial dimensions, will ignore the longer part, if shorter, will pad with 1.0. if the components of the pixdim are non-positive values, the transform will use the corresponding components of the original pixdim, which is computed from the affine matrix of input image.

  • 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

  • padding_mode (Union[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

  • 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 (Union[dtype, type, 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.

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.

See Also: nibabel.orientations.ornt2axcodes.

__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.

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

See Also: nibabel.orientations.ornt2axcodes.

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, padding_mode=GridSamplePadMode.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

  • padding_mode (Union[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

  • align_corners (bool) – Defaults to False. See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample

  • dtype (Union[dtype, type, None, 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__(img, mode=None, padding_mode=None, align_corners=None, dtype=None)[source]
Parameters
  • img (Union[ndarray, Tensor]) – channel first array, must have shape 2D: (nchannels, H, W), or 3D: (nchannels, 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

  • padding_mode (Union[GridSamplePadMode, str, None]) – {"zeros", "border", "reflection"} Padding mode for outside grid values. Defaults to self.padding_mode. See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample

  • align_corners (Optional[bool]) – Defaults to self.align_corners. See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample

  • dtype (Union[dtype, type, None, 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.

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.

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

Return type

Union[ndarray, Tensor]

RandAxisFlip

class monai.transforms.RandAxisFlip(prob=0.1)[source]

Randomly select a spatial axis and flip along it. See numpy.flip for additional details. https://docs.scipy.org/doc/numpy/reference/generated/numpy.flip.html

Parameters

prob (float) – Probability of flipping.

__call__(img)[source]
Parameters

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

Return type

Union[ndarray, Tensor]

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.

Return type

None

RandZoom

class monai.transforms.RandZoom(prob=0.1, min_zoom=0.9, max_zoom=1.1, mode=InterpolateMode.AREA, padding_mode=NumpyPadMode.EDGE, align_corners=None, keep_size=True, **kwargs)[source]

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, PytorchPadMode, str]) – available modes for numpy array:{"constant", "edge", "linear_ramp", "maximum", "mean", "median", "minimum", "reflect", "symmetric", "wrap", "empty"} available modes for PyTorch Tensor: {"constant", "reflect", "replicate", "circular"}. One of the listed string values or a user supplied function. Defaults to "constant". The mode to pad data after zooming. See also: https://numpy.org/doc/1.18/reference/generated/numpy.pad.html https://pytorch.org/docs/stable/generated/torch.nn.functional.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.

  • kwargs – other arguments for the np.pad or torch.pad function. note that np.pad treats channel dimension as the first dimension.

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

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, padding_mode=GridSamplePadMode.REFLECTION, as_tensor_output=True, device=None, image_only=False)[source]

Transform img given the affine parameters. A tutorial is available: https://github.com/Project-MONAI/tutorials/blob/0.6.0/modules/transforms_demo_2d.ipynb.

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]) –

    shearing factors for affine matrix, take a 3D affine as example:

    [
        [1.0, params[0], params[1], 0.0],
        [params[2], 1.0, params[3], 0.0],
        [params[4], params[5], 1.0, 0.0],
        [0.0, 0.0, 0.0, 1.0],
    ]
    
    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]) – scale factor for every spatial dims. a tuple of 2 floats for 2D, a tuple of 3 floats for 3D. Defaults to 1.0.

  • 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 some 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

  • padding_mode (Union[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

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

  • image_only (bool) – if True return only the image volume, otherwise return (image, affine).

__call__(img, spatial_size=None, mode=None, padding_mode=None)[source]
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

  • padding_mode (Union[GridSamplePadMode, str, None]) – {"zeros", "border", "reflection"} Padding mode for outside grid values. Defaults to self.padding_mode. See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample

Return type

Union[ndarray, Tensor, Tuple[Union[ndarray, Tensor], Union[ndarray, Tensor]]]

__init__(rotate_params=None, shear_params=None, translate_params=None, scale_params=None, spatial_size=None, mode=GridSampleMode.BILINEAR, padding_mode=GridSamplePadMode.REFLECTION, as_tensor_output=True, device=None, image_only=False)[source]

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]) –

    shearing factors for affine matrix, take a 3D affine as example:

    [
        [1.0, params[0], params[1], 0.0],
        [params[2], 1.0, params[3], 0.0],
        [params[4], params[5], 1.0, 0.0],
        [0.0, 0.0, 0.0, 1.0],
    ]
    
    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]) – scale factor for every spatial dims. a tuple of 2 floats for 2D, a tuple of 3 floats for 3D. Defaults to 1.0.

  • 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 some 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

  • padding_mode (Union[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

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

  • image_only (bool) – if True return only the image volume, otherwise return (image, affine).

Resample

class monai.transforms.Resample(mode=GridSampleMode.BILINEAR, padding_mode=GridSamplePadMode.BORDER, as_tensor_output=True, device=None)[source]

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

Parameters
__call__(img, grid=None, mode=None, padding_mode=None)[source]
Parameters
Return type

Union[ndarray, Tensor]

__init__(mode=GridSampleMode.BILINEAR, padding_mode=GridSamplePadMode.BORDER, as_tensor_output=True, device=None)[source]

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

Parameters

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, padding_mode=GridSamplePadMode.REFLECTION, cache_grid=False, as_tensor_output=True, device=None)[source]

Random affine transform. A tutorial is available: https://github.com/Project-MONAI/tutorials/blob/0.6.0/modules/transforms_demo_2d.ipynb.

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[Union[Tuple[float, float], float]], float, None]) – angle range in radians. If element i is a pair of (min, max) values, then uniform[-rotate_range[i][0], rotate_range[i][1]) will be used to generate the rotation parameter for the i`th spatial dimension. If not, `uniform[-rotate_range[i], rotate_range[i]) will be used. This can be altered on a per-dimension basis. E.g., ((0,3), 1, …): for dim0, rotation will be in range [0, 3], and for dim1 [-1, 1] will be used. Setting a single value will use [-x, x] for dim0 and nothing for the remaining dimensions.

  • shear_range (Union[Sequence[Union[Tuple[float, float], float]], float, None]) –

    shear range with format matching rotate_range, it defines the range to randomly select shearing factors(a tuple of 2 floats for 2D, a tuple of 6 floats for 3D) for affine matrix, take a 3D affine as example:

    [
        [1.0, params[0], params[1], 0.0],
        [params[2], 1.0, params[3], 0.0],
        [params[4], params[5], 1.0, 0.0],
        [0.0, 0.0, 0.0, 1.0],
    ]
    

  • translate_range (Union[Sequence[Union[Tuple[float, float], float]], float, None]) – translate range with format matching rotate_range, it defines the range to randomly select pixel/voxel to translate for every spatial dims.

  • scale_range (Union[Sequence[Union[Tuple[float, float], float]], float, None]) – scaling range with format matching rotate_range. it defines the range to randomly select the scale factor to translate for every spatial dims. A value of 1.0 is added to the result. This allows 0 to correspond to no change (i.e., a scaling of 1.0).

  • 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 some 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

  • padding_mode (Union[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

  • cache_grid (bool) – whether to cache the identity sampling grid. If the spatial size is not dynamically defined by input image, enabling this option could accelerate the transform.

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

See also

  • RandAffineGrid for the random affine parameters configurations.

  • Affine for the affine transformation parameters configurations.

__call__(img, spatial_size=None, mode=None, padding_mode=None)[source]
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

  • padding_mode (Union[GridSamplePadMode, str, None]) – {"zeros", "border", "reflection"} Padding mode for outside grid values. Defaults to self.padding_mode. See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample

Return type

Union[ndarray, Tensor]

__init__(prob=0.1, rotate_range=None, shear_range=None, translate_range=None, scale_range=None, spatial_size=None, mode=GridSampleMode.BILINEAR, padding_mode=GridSamplePadMode.REFLECTION, cache_grid=False, as_tensor_output=True, device=None)[source]
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[Union[Tuple[float, float], float]], float, None]) – angle range in radians. If element i is a pair of (min, max) values, then uniform[-rotate_range[i][0], rotate_range[i][1]) will be used to generate the rotation parameter for the i`th spatial dimension. If not, `uniform[-rotate_range[i], rotate_range[i]) will be used. This can be altered on a per-dimension basis. E.g., ((0,3), 1, …): for dim0, rotation will be in range [0, 3], and for dim1 [-1, 1] will be used. Setting a single value will use [-x, x] for dim0 and nothing for the remaining dimensions.

  • shear_range (Union[Sequence[Union[Tuple[float, float], float]], float, None]) –

    shear range with format matching rotate_range, it defines the range to randomly select shearing factors(a tuple of 2 floats for 2D, a tuple of 6 floats for 3D) for affine matrix, take a 3D affine as example:

    [
        [1.0, params[0], params[1], 0.0],
        [params[2], 1.0, params[3], 0.0],
        [params[4], params[5], 1.0, 0.0],
        [0.0, 0.0, 0.0, 1.0],
    ]
    

  • translate_range (Union[Sequence[Union[Tuple[float, float], float]], float, None]) – translate range with format matching rotate_range, it defines the range to randomly select pixel/voxel to translate for every spatial dims.

  • scale_range (Union[Sequence[Union[Tuple[float, float], float]], float, None]) – scaling range with format matching rotate_range. it defines the range to randomly select the scale factor to translate for every spatial dims. A value of 1.0 is added to the result. This allows 0 to correspond to no change (i.e., a scaling of 1.0).

  • 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 some 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

  • padding_mode (Union[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

  • cache_grid (bool) – whether to cache the identity sampling grid. If the spatial size is not dynamically defined by input image, enabling this option could accelerate the transform.

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

See also

  • RandAffineGrid for the random affine parameters configurations.

  • Affine for the affine transformation parameters configurations.

get_identity_grid(spatial_size)[source]

Return a cached or new identity grid depends on the availability.

Parameters

spatial_size (Sequence[int]) – non-dynamic spatial size

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.

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.

__init__(spacing, magnitude_range, as_tensor_output=True, device=None)[source]
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.

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, affine=None)[source]

Affine transforms on the coordinates.

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]) –

    shearing factors for affine matrix, take a 3D affine as example:

    [
        [1.0, params[0], params[1], 0.0],
        [params[2], 1.0, params[3], 0.0],
        [params[4], params[5], 1.0, 0.0],
        [0.0, 0.0, 0.0, 1.0],
    ]
    
    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]) – scale factor for every spatial dims. a tuple of 2 floats for 2D, a tuple of 3 floats for 3D. Defaults to 1.0.

  • affine (Union[ndarray, Tensor, None]) – If applied, ignore the params (rotate_params, etc.) and use the supplied matrix. Should be square with each side = num of image spatial dimensions + 1.

__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

Tuple[Union[ndarray, Tensor], 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[Union[Tuple[float, float], float]], float, None]) – angle range in radians. If element i is a pair of (min, max) values, then uniform[-rotate_range[i][0], rotate_range[i][1]) will be used to generate the rotation parameter for the i`th spatial dimension. If not, `uniform[-rotate_range[i], rotate_range[i]) will be used. This can be altered on a per-dimension basis. E.g., ((0,3), 1, …): for dim0, rotation will be in range [0, 3], and for dim1 [-1, 1] will be used. Setting a single value will use [-x, x] for dim0 and nothing for the remaining dimensions.

  • shear_range (Union[Sequence[Union[Tuple[float, float], float]], float, None]) –

    shear range with format matching rotate_range, it defines the range to randomly select shearing factors(a tuple of 2 floats for 2D, a tuple of 6 floats for 3D) for affine matrix, take a 3D affine as example:

    [
        [1.0, params[0], params[1], 0.0],
        [params[2], 1.0, params[3], 0.0],
        [params[4], params[5], 1.0, 0.0],
        [0.0, 0.0, 0.0, 1.0],
    ]
    

  • translate_range (Union[Sequence[Union[Tuple[float, float], float]], float, None]) – translate range with format matching rotate_range, it defines the range to randomly select voxels to translate for every spatial dims.

  • scale_range (Union[Sequence[Union[Tuple[float, float], float]], float, None]) – scaling range with format matching rotate_range. it defines the range to randomly select the scale factor to translate for every spatial dims. A value of 1.0 is added to the result. This allows 0 to correspond to no change (i.e., a scaling of 1.0).

  • 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.

__init__(rotate_range=None, shear_range=None, translate_range=None, scale_range=None, as_tensor_output=True, device=None)[source]
Parameters
  • rotate_range (Union[Sequence[Union[Tuple[float, float], float]], float, None]) – angle range in radians. If element i is a pair of (min, max) values, then uniform[-rotate_range[i][0], rotate_range[i][1]) will be used to generate the rotation parameter for the i`th spatial dimension. If not, `uniform[-rotate_range[i], rotate_range[i]) will be used. This can be altered on a per-dimension basis. E.g., ((0,3), 1, …): for dim0, rotation will be in range [0, 3], and for dim1 [-1, 1] will be used. Setting a single value will use [-x, x] for dim0 and nothing for the remaining dimensions.

  • shear_range (Union[Sequence[Union[Tuple[float, float], float]], float, None]) –

    shear range with format matching rotate_range, it defines the range to randomly select shearing factors(a tuple of 2 floats for 2D, a tuple of 6 floats for 3D) for affine matrix, take a 3D affine as example:

    [
        [1.0, params[0], params[1], 0.0],
        [params[2], 1.0, params[3], 0.0],
        [params[4], params[5], 1.0, 0.0],
        [0.0, 0.0, 0.0, 1.0],
    ]
    

  • translate_range (Union[Sequence[Union[Tuple[float, float], float]], float, None]) – translate range with format matching rotate_range, it defines the range to randomly select voxels to translate for every spatial dims.

  • scale_range (Union[Sequence[Union[Tuple[float, float], float]], float, None]) – scaling range with format matching rotate_range. it defines the range to randomly select the scale factor to translate for every spatial dims. A value of 1.0 is added to the result. This allows 0 to correspond to no change (i.e., a scaling of 1.0).

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

get_transformation_matrix()[source]

Get the most recently applied transformation matrix

Return type

Union[ndarray, Tensor, None]

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, padding_mode=GridSamplePadMode.REFLECTION, as_tensor_output=False, device=None)[source]

Random elastic deformation and affine in 2D. A tutorial is available: https://github.com/Project-MONAI/tutorials/blob/0.6.0/modules/transforms_demo_2d.ipynb.

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[Union[Tuple[float, float], float]], float, None]) – angle range in radians. If element i is a pair of (min, max) values, then uniform[-rotate_range[i][0], rotate_range[i][1]) will be used to generate the rotation parameter for the i`th spatial dimension. If not, `uniform[-rotate_range[i], rotate_range[i]) will be used. This can be altered on a per-dimension basis. E.g., ((0,3), 1, …): for dim0, rotation will be in range [0, 3], and for dim1 [-1, 1] will be used. Setting a single value will use [-x, x] for dim0 and nothing for the remaining dimensions.

  • shear_range (Union[Sequence[Union[Tuple[float, float], float]], float, None]) –

    shear range with format matching rotate_range, it defines the range to randomly select shearing factors(a tuple of 2 floats for 2D) for affine matrix, take a 2D affine as example:

    [
        [1.0, params[0], 0.0],
        [params[1], 1.0, 0.0],
        [0.0, 0.0, 1.0],
    ]
    

  • translate_range (Union[Sequence[Union[Tuple[float, float], float]], float, None]) – translate range with format matching rotate_range, it defines the range to randomly select pixel to translate for every spatial dims.

  • scale_range (Union[Sequence[Union[Tuple[float, float], float]], float, None]) – scaling range with format matching rotate_range. it defines the range to randomly select the scale factor to translate for every spatial dims. A value of 1.0 is added to the result. This allows 0 to correspond to no change (i.e., a scaling of 1.0).

  • spatial_size (Union[int, Tuple[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 some 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

  • padding_mode (Union[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

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

See also

  • RandAffineGrid for the random affine parameters configurations.

  • Affine for the affine transformation parameters configurations.

__call__(img, spatial_size=None, mode=None, padding_mode=None)[source]
Parameters
  • img (Union[ndarray, Tensor]) – shape must be (num_channels, H, W),

  • spatial_size (Union[int, Tuple[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

  • padding_mode (Union[GridSamplePadMode, str, None]) – {"zeros", "border", "reflection"} Padding mode for outside grid values. Defaults to self.padding_mode. See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample

Return type

Union[ndarray, Tensor]

__init__(spacing, magnitude_range, prob=0.1, rotate_range=None, shear_range=None, translate_range=None, scale_range=None, spatial_size=None, mode=GridSampleMode.BILINEAR, padding_mode=GridSamplePadMode.REFLECTION, as_tensor_output=False, device=None)[source]
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[Union[Tuple[float, float], float]], float, None]) – angle range in radians. If element i is a pair of (min, max) values, then uniform[-rotate_range[i][0], rotate_range[i][1]) will be used to generate the rotation parameter for the i`th spatial dimension. If not, `uniform[-rotate_range[i], rotate_range[i]) will be used. This can be altered on a per-dimension basis. E.g., ((0,3), 1, …): for dim0, rotation will be in range [0, 3], and for dim1 [-1, 1] will be used. Setting a single value will use [-x, x] for dim0 and nothing for the remaining dimensions.

  • shear_range (Union[Sequence[Union[Tuple[float, float], float]], float, None]) –

    shear range with format matching rotate_range, it defines the range to randomly select shearing factors(a tuple of 2 floats for 2D) for affine matrix, take a 2D affine as example:

    [
        [1.0, params[0], 0.0],
        [params[1], 1.0, 0.0],
        [0.0, 0.0, 1.0],
    ]
    

  • translate_range (Union[Sequence[Union[Tuple[float, float], float]], float, None]) – translate range with format matching rotate_range, it defines the range to randomly select pixel to translate for every spatial dims.

  • scale_range (Union[Sequence[Union[Tuple[float, float], float]], float, None]) – scaling range with format matching rotate_range. it defines the range to randomly select the scale factor to translate for every spatial dims. A value of 1.0 is added to the result. This allows 0 to correspond to no change (i.e., a scaling of 1.0).

  • spatial_size (Union[int, Tuple[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 some 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

  • padding_mode (Union[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

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

See also

  • RandAffineGrid for the random affine parameters configurations.

  • Affine for the affine transformation parameters configurations.

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.

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, padding_mode=GridSamplePadMode.REFLECTION, as_tensor_output=False, device=None)[source]

Random elastic deformation and affine in 3D. A tutorial is available: https://github.com/Project-MONAI/tutorials/blob/0.6.0/modules/transforms_demo_2d.ipynb.

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[Union[Tuple[float, float], float]], float, None]) – angle range in radians. If element i is a pair of (min, max) values, then uniform[-rotate_range[i][0], rotate_range[i][1]) will be used to generate the rotation parameter for the i`th spatial dimension. If not, `uniform[-rotate_range[i], rotate_range[i]) will be used. This can be altered on a per-dimension basis. E.g., ((0,3), 1, …): for dim0, rotation will be in range [0, 3], and for dim1 [-1, 1] will be used. Setting a single value will use [-x, x] for dim0 and nothing for the remaining dimensions.

  • shear_range (Union[Sequence[Union[Tuple[float, float], float]], float, None]) –

    shear range with format matching rotate_range, it defines the range to randomly select shearing factors(a tuple of 6 floats for 3D) for affine matrix, take a 3D affine as example:

    [
        [1.0, params[0], params[1], 0.0],
        [params[2], 1.0, params[3], 0.0],
        [params[4], params[5], 1.0, 0.0],
        [0.0, 0.0, 0.0, 1.0],
    ]
    

  • translate_range (Union[Sequence[Union[Tuple[float, float], float]], float, None]) – translate range with format matching rotate_range, it defines the range to randomly select voxel to translate for every spatial dims.

  • scale_range (Union[Sequence[Union[Tuple[float, float], float]], float, None]) – scaling range with format matching rotate_range. it defines the range to randomly select the scale factor to translate for every spatial dims. A value of 1.0 is added to the result. This allows 0 to correspond to no change (i.e., a scaling of 1.0).

  • spatial_size (Union[Tuple[int, int, 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 some 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

  • padding_mode (Union[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

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

See also

  • RandAffineGrid for the random affine parameters configurations.

  • Affine for the affine transformation parameters configurations.

__call__(img, spatial_size=None, mode=None, padding_mode=None)[source]
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

  • padding_mode (Union[GridSamplePadMode, str, None]) – {"zeros", "border", "reflection"} Padding mode for outside grid values. Defaults to self.padding_mode. See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample

Return type

Union[ndarray, Tensor]

__init__(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, padding_mode=GridSamplePadMode.REFLECTION, as_tensor_output=False, device=None)[source]
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[Union[Tuple[float, float], float]], float, None]) – angle range in radians. If element i is a pair of (min, max) values, then uniform[-rotate_range[i][0], rotate_range[i][1]) will be used to generate the rotation parameter for the i`th spatial dimension. If not, `uniform[-rotate_range[i], rotate_range[i]) will be used. This can be altered on a per-dimension basis. E.g., ((0,3), 1, …): for dim0, rotation will be in range [0, 3], and for dim1 [-1, 1] will be used. Setting a single value will use [-x, x] for dim0 and nothing for the remaining dimensions.

  • shear_range (Union[Sequence[Union[Tuple[float, float], float]], float, None]) –

    shear range with format matching rotate_range, it defines the range to randomly select shearing factors(a tuple of 6 floats for 3D) for affine matrix, take a 3D affine as example:

    [
        [1.0, params[0], params[1], 0.0],
        [params[2], 1.0, params[3], 0.0],
        [params[4], params[5], 1.0, 0.0],
        [0.0, 0.0, 0.0, 1.0],
    ]
    

  • translate_range (Union[Sequence[Union[Tuple[float, float], float]], float, None]) – translate range with format matching rotate_range, it defines the range to randomly select voxel to translate for every spatial dims.

  • scale_range (Union[Sequence[Union[Tuple[float, float], float]], float, None]) – scaling range with format matching rotate_range. it defines the range to randomly select the scale factor to translate for every spatial dims. A value of 1.0 is added to the result. This allows 0 to correspond to no change (i.e., a scaling of 1.0).

  • spatial_size (Union[Tuple[int, int, 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 some 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

  • padding_mode (Union[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

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

See also

  • RandAffineGrid for the random affine parameters configurations.

  • Affine for the affine transformation parameters configurations.

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.

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. See np.rot90 for additional details: https://numpy.org/doc/stable/reference/generated/numpy.rot90.html.

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. If axis is negative it counts from the last to the first axis.

__call__(img)[source]
Parameters

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

Return type

Union[ndarray, Tensor]

__init__(k=1, spatial_axes=(0, 1))[source]
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. If axis is negative it counts from the last to the first axis.

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

Return type

Union[ndarray, Tensor]

__init__(prob=0.1, max_k=3, spatial_axes=(0, 1))[source]
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.

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.

Return type

None

Flip

class monai.transforms.Flip(spatial_axis=None)[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. The default axis=None will flip over all of the axes of the input array. If axis is negative it counts from the last to the first axis. If axis is a tuple of ints, flipping is performed on all of the axes specified in the tuple.

__call__(img)[source]
Parameters

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

Return type

Union[ndarray, Tensor]

Resize

class monai.transforms.Resize(spatial_size, size_mode='all', mode=InterpolateMode.AREA, align_corners=None)[source]

Resize the input image to given spatial size (with scaling, not cropping/padding). Implemented using torch.nn.functional.interpolate.

Parameters
  • spatial_size (Union[Sequence[int], int]) – expected shape of spatial dimensions after resize operation. if some 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.

  • size_mode (str) – should be “all” or “longest”, if “all”, will use spatial_size for all the spatial dims, if “longest”, rescale the image so that only the longest side is equal to specified spatial_size, which must be an int number in this case, keeping the aspect ratio of the initial image, refer to: https://albumentations.ai/docs/api_reference/augmentations/geometric/resize/ #albumentations.augmentations.geometric.resize.LongestMaxSize.

  • 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

class monai.transforms.Rotate(angle, keep_size=True, mode=GridSampleMode.BILINEAR, padding_mode=GridSamplePadMode.BORDER, align_corners=False, dtype=<class 'numpy.float64'>)[source]

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

  • padding_mode (Union[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

  • align_corners (bool) – Defaults to False. See also: https://pytorch.org/docs/stable/nn.functional.html#grid-sample

  • dtype (Union[dtype, type, None, 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__(img, mode=None, padding_mode=None, align_corners=None, dtype=None)[source]
Parameters
Raises

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

Return type

Union[ndarray, Tensor]

get_rotation_matrix()[source]

Get the most recently applied rotation matrix This is not thread-safe.

Return type

Optional[ndarray]

Zoom

class monai.transforms.Zoom(zoom, mode=InterpolateMode.AREA, padding_mode=NumpyPadMode.EDGE, align_corners=None, keep_size=True, **kwargs)[source]

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, PytorchPadMode, str]) – available modes for numpy array:{"constant", "edge", "linear_ramp", "maximum", "mean", "median", "minimum", "reflect", "symmetric", "wrap", "empty"} available modes for PyTorch Tensor: {"constant", "reflect", "replicate", "circular"}. One of the listed string values or a user supplied function. Defaults to "constant". The mode to pad data after zooming. See also: https://numpy.org/doc/1.18/reference/generated/numpy.pad.html https://pytorch.org/docs/stable/generated/torch.nn.functional.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.

  • kwargs – other arguments for the np.pad or torch.pad function. note that np.pad treats channel dimension as the first dimension.

__call__(img, mode=None, padding_mode=None, align_corners=None)[source]
Parameters
Return type

Union[ndarray, Tensor]

AddCoordinateChannels

class monai.transforms.AddCoordinateChannels(spatial_channels)[source]

Appends additional channels encoding coordinates of the input. Useful when e.g. training using patch-based sampling, to allow feeding of the patch’s location into the network.

This can be seen as a input-only version of CoordConv:

Liu, R. et al. An Intriguing Failing of Convolutional Neural Networks and the CoordConv Solution, NeurIPS 2018.

Parameters

spatial_channels (Sequence[int]) – the spatial dimensions that are to have their coordinates encoded in a channel and appended to the input. E.g., (1,2,3) will append three channels to the input, encoding the coordinates of the input’s three spatial dimensions (0 is reserved for the channel dimension).

__call__(img)[source]
Parameters

img (Union[ndarray, Tensor]) – data to be transformed, assuming img is channel first.

__init__(spatial_channels)[source]
Parameters

spatial_channels (Sequence[int]) – the spatial dimensions that are to have their coordinates encoded in a channel and appended to the input. E.g., (1,2,3) will append three channels to the input, encoding the coordinates of the input’s three spatial dimensions (0 is reserved for the channel dimension).

Utility

Identity

class monai.transforms.Identity[source]

Do nothing to the 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

Union[ndarray, Tensor]

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

Union[ndarray, Tensor]

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

Union[ndarray, Tensor]

AddChannel

class monai.transforms.AddChannel[source]

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

Union[ndarray, Tensor]

EnsureChannelFirst

class monai.transforms.EnsureChannelFirst(strict_check=True)[source]

Automatically adjust or add the channel dimension of input data to ensure channel_first shape. It extracts the original_channel_dim info from provided meta_data dictionary. Typical values of original_channel_dim can be: “no_channel”, 0, -1. Convert the data to channel_first based on the original_channel_dim information.

Parameters

strict_check (bool) – whether to raise an error when the meta information is insufficient.

__call__(img, meta_dict=None)[source]

Apply the transform to img.

Return type

Union[ndarray, Tensor]

__init__(strict_check=True)[source]
Parameters

strict_check (bool) – whether to raise an error when the meta information is insufficient.

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

Union[ndarray, Tensor]

SplitChannel

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

Split Numpy array or PyTorch Tensor data according to the channel dim. It can help applying different following transforms to different channels.

Parameters

channel_dim (int) – which dimension of input image is the channel, default to 0.

__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 – 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, type, None, dtype]) – 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]

__init__(dtype=<class 'numpy.float32'>)[source]
Parameters

dtype – convert image to this data type, default is np.float32.

ToTensor

class monai.transforms.ToTensor(dtype=None, device=None)[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(dtype=None)[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

ToCupy

class monai.transforms.ToCupy(dtype=None)[source]

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

Parameters

dtype – data type specifier. It is inferred from the input by default.

__call__(data)[source]

Create a CuPy array from data and make it contiguous

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

Union[ndarray, Tensor]

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 (Union[ndarray, Tensor]) – numpy arrays with required dimension dim removed

Return type

Union[ndarray, Tensor]

__init__(dim=0)[source]
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].

DataStats

class monai.transforms.DataStats(prefix='Data', data_type=True, 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_type (bool) – whether to show the type of input data.

  • 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 the handler should have a logging level of at least INFO.

Raises

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

__call__(img, prefix=None, data_type=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

Union[ndarray, Tensor]

__init__(prefix='Data', data_type=True, data_shape=True, value_range=True, data_value=False, additional_info=None, logger_handler=None)[source]
Parameters
  • prefix (str) – will be printed in format: “{prefix} statistics”.

  • data_type (bool) – whether to show the type of input data.

  • 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 the handler should have a logging level of at least INFO.

Raises

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

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 (Union[ndarray, Tensor]) – 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

Union[ndarray, Tensor]

__init__(delay_time=0.0)[source]
Parameters

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

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.

RandLambda

class monai.transforms.RandLambda(func=None, prob=1.0)[source]

Randomizable version monai.transforms.Lambda, the input func may contain random logic, or randomly execute the function based on prob.

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

  • prob (float) – probability of executing the random function, default to 1.0, with 100% probability to execute.

For more details, please check monai.transforms.Lambda.

__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.

LabelToMask

class monai.transforms.LabelToMask(select_labels, merge_channels=False)[source]

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

Union[ndarray, Tensor]

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]

ClassesToIndices

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

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

Parameters
  • num_classes (Optional[int]) – number of classes for argmax label, not necessary for One-Hot label.

  • image_threshold (float) – if enabled image at runtime, use image > image_threshold to determine the valid image content area and select only the indices of classes 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 the indices of every class.

  • image (Optional[ndarray]) – if image is not None, use image > image_threshold to define valid region, and only select the indices within the valid region.

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

Return type

List[ndarray]

__init__(num_classes=None, image_threshold=0.0, output_shape=None)[source]

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

Parameters
  • num_classes (Optional[int]) – number of classes for argmax label, not necessary for One-Hot label.

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

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

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

AddExtremePointsChannel

class monai.transforms.AddExtremePointsChannel(background=0, pert=0.0)[source]

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.

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.

__init__(name, *args, **kwargs)[source]
Parameters
  • name (str) – The transform name in TorchVision package.

  • args – parameters for the TorchVision transform.

  • kwargs – parameters for the TorchVision transform.

MapLabelValue

class monai.transforms.MapLabelValue(orig_labels, target_labels, dtype=<class 'numpy.float32'>)[source]

Utility to map label values to another set of values. For example, map [3, 2, 1] to [0, 1, 2], [1, 2, 3] -> [0.5, 1.5, 2.5], [“label3”, “label2”, “label1”] -> [0, 1, 2], [3.5, 2.5, 1.5] -> [“label0”, “label1”, “label2”], etc. The label data must be numpy array or array-like data and the output data will be numpy array.

Parameters
  • orig_labels (Sequence) – original labels that map to others.

  • target_labels (Sequence) – expected label values, 1: 1 map to the orig_labels.

  • dtype (Union[dtype, type, None]) – convert the output data to dtype, default to float32.

__call__(img)[source]

Call self as a function.

__init__(orig_labels, target_labels, dtype=<class 'numpy.float32'>)[source]
Parameters
  • orig_labels (Sequence) – original labels that map to others.

  • target_labels (Sequence) – expected label values, 1: 1 map to the orig_labels.

  • dtype (Union[dtype, type, None]) – convert the output data to dtype, default to float32.

EnsureType

class monai.transforms.EnsureType(data_type='tensor', dtype=None, device=None)[source]

Ensure the input data to be a PyTorch Tensor or numpy array, support: numpy array, PyTorch Tensor, float, int, bool, string and object keep the original. If passing a dictionary, list or tuple, still return dictionary, list or tuple and recursively convert every item to the expected data type.

Parameters
  • data_type (str) – target data type to convert, should be “tensor” or “numpy”.

  • dtype (Union[dtype, type, None, dtype]) – target data content type to convert, for example: np.float32, torch.float, etc.

  • device (Optional[device]) – for Tensor data type, specify the target device.

__call__(data)[source]
Parameters

data (Union[ndarray, Tensor]) – input data can be PyTorch Tensor, numpy array, list, dictionary, int, float, bool, str, etc. will ensure Tensor, Numpy array, float, int, bool as Tensors or numpy arrays, strings and objects keep the original. for dictionary, list or tuple, ensure every item as expected type if applicable.

IntensityStats

class monai.transforms.IntensityStats(ops, key_prefix, channel_wise=False)[source]

Compute statistics for the intensity values of input image and store into the meta data dictionary. For example: if ops=[lambda x: np.mean(x), “max”] and key_prefix=”orig”, may generate below stats: {“orig_custom_0”: 1.5, “orig_max”: 3.0}.

Parameters
  • ops (Sequence[Union[str, Callable]]) – expected operations to compute statistics for the intensity. if a string, will map to the predefined operations, supported: [“mean”, “median”, “max”, “min”, “std”] mapping to np.nanmean, np.nanmedian, np.nanmax, np.nanmin, np.nanstd. if a callable function, will execute the function on input image.

  • key_prefix (str) – the prefix to combine with ops name to generate the key to store the results in the meta data dictionary. if some ops are callable functions, will use “{key_prefix}_custom_{index}” as the key, where index counts from 0.

  • channel_wise (bool) – whether to compute statistics for every channel of input image separately. if True, return a list of values for every operation, default to False.

__call__(img, meta_data=None, mask=None)[source]

Compute statistics for the intensity of input image.

Parameters
  • img (ndarray) – input image to compute intensity stats.

  • meta_data (Optional[Dict]) – meta data dictionary to store the statistics data, if None, will create an empty dictionary.

  • mask (Optional[ndarray]) – if not None, mask the image to extract only the interested area to compute statistics. mask must have the same shape as input img.

Return type

Tuple[ndarray, Dict]

ToDevice

class monai.transforms.ToDevice(device, **kwargs)[source]

Move PyTorch Tensor to the specified device. It can help cache data into GPU and execute following logic on GPU directly.

Note

If moving data to GPU device in the multi-processing workers of DataLoader, may got below CUDA error: “RuntimeError: Cannot re-initialize CUDA in forked subprocess. To use CUDA with multiprocessing, you must use the ‘spawn’ start method.” So usually suggest to set num_workers=0 in the DataLoader or ThreadDataLoader.

Parameters
__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.

__init__(device, **kwargs)[source]
Parameters

CuCIM

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

Wrap a non-randomized cuCIM transform, defined based on the transform name and args. For randomized transforms (or randomly applying a transform) use monai.transforms.RandCuCIM.

Parameters
  • name (str) – the transform name in CuCIM package

  • args – parameters for the CuCIM transform

  • kwargs – parameters for the CuCIM transform

Note

CuCIM transform only work with CuPy arrays, so this transform expects input data to be cupy.ndarray. Users can call ToCuPy transform to convert a numpy array or torch tensor to cupy array.

__call__(data)[source]
Parameters

data – a CuPy array (cupy.ndarray) for the cuCIM transform

Returns

cupy.ndarray

RandCuCIM

class monai.transforms.RandCuCIM(name, apply_prob=1.0, *args, **kwargs)[source]

Wrap a randomized cuCIM transform, defined based on the transform name and args, or randomly apply a non-randomized transform. For deterministic non-randomized transforms use monai.transforms.CuCIM.

Parameters
  • name (str) – the transform name in CuCIM package.

  • apply_prob (float) – the probability to apply the transform (default=1.0)

  • args – parameters for the CuCIM transform.

  • kwargs – parameters for the CuCIM transform.

Note

  • CuCIM transform only work with CuPy arrays, so this transform expects input data to be cupy.ndarray. Users can call ToCuPy transform to convert a numpy array or torch tensor to cupy array.

  • If the cuCIM transform is already randomized the apply_prob argument has nothing to do with the randomness of the underlying cuCIM transform. apply_prob defines if the transform (either randomized or non-randomized) being applied randomly, so it can apply non-randomized tranforms randomly but be careful with setting apply_prob to anything than 1.0 when using along with cuCIM’s randomized transforms.

  • If the random factor of the underlying cuCIM transform is not derived from self.R, the results may not be deterministic. See Also: monai.transforms.Randomizable.

__call__(data)[source]
Parameters

data – a CuPy array (cupy.ndarray) for the cuCIM transform

Returns

cupy.ndarray

Dictionary Transforms

Crop and Pad (Dict)

SpatialPadd

class monai.transforms.SpatialPadd(keys, spatial_size, method=Method.SYMMETRIC, mode=NumpyPadMode.CONSTANT, allow_missing_keys=False, **kwargs)[source]

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 a dimension of the input data size is bigger than the pad size, will not pad that dimension. If its components have non-positive values, the corresponding size of input image will be used. for example: if the spatial size of input data is [30, 30, 30] and spatial_size=[32, 25, -1], the spatial size of output data will be [32, 30, 30].

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

  • mode (Union[Sequence[Union[NumpyPadMode, PytorchPadMode, str]], NumpyPadMode, PytorchPadMode, str]) – available modes for numpy array:{"constant", "edge", "linear_ramp", "maximum", "mean", "median", "minimum", "reflect", "symmetric", "wrap", "empty"} available modes for PyTorch Tensor: {"constant", "reflect", "replicate", "circular"}. 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 https://pytorch.org/docs/stable/generated/torch.nn.functional.pad.html It also can be a sequence of string, each element corresponds to a key in keys.

  • allow_missing_keys (bool) – don’t raise exception if key is missing.

  • kwargs – other arguments for the np.pad or torch.pad function. note that np.pad treats channel dimension as 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.

__init__(keys, spatial_size, method=Method.SYMMETRIC, mode=NumpyPadMode.CONSTANT, allow_missing_keys=False, **kwargs)[source]
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 a dimension of the input data size is bigger than the pad size, will not pad that dimension. If its components have non-positive values, the corresponding size of input image will be used. for example: if the spatial size of input data is [30, 30, 30] and spatial_size=[32, 25, -1], the spatial size of output data will be [32, 30, 30].

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

  • mode (Union[Sequence[Union[NumpyPadMode, PytorchPadMode, str]], NumpyPadMode, PytorchPadMode, str]) – available modes for numpy array:{"constant", "edge", "linear_ramp", "maximum", "mean", "median", "minimum", "reflect", "symmetric", "wrap", "empty"} available modes for PyTorch Tensor: {"constant", "reflect", "replicate", "circular"}. 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 https://pytorch.org/docs/stable/generated/torch.nn.functional.pad.html It also can be a sequence of string, each element corresponds to a key in keys.

  • allow_missing_keys (bool) – don’t raise exception if key is missing.

  • kwargs – other arguments for the np.pad or torch.pad function. note that np.pad treats channel dimension as the first dimension.

inverse(data)[source]

Inverse of __call__.

Raises

NotImplementedError – When the subclass does not override this method.

Return type

Dict[Hashable, Union[ndarray, Tensor]]

BorderPadd

class monai.transforms.BorderPadd(keys, spatial_border, mode=NumpyPadMode.CONSTANT, allow_missing_keys=False, **kwargs)[source]

Pad the input data by adding specified borders to every dimension. Dictionary-based wrapper of monai.transforms.BorderPad.

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, PytorchPadMode, str]], NumpyPadMode, PytorchPadMode, str]) – available modes for numpy array:{"constant", "edge", "linear_ramp", "maximum", "mean", "median", "minimum", "reflect", "symmetric", "wrap", "empty"} available modes for PyTorch Tensor: {"constant", "reflect", "replicate", "circular"}. 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 https://pytorch.org/docs/stable/generated/torch.nn.functional.pad.html It also can be a sequence of string, each element corresponds to a key in keys.

  • allow_missing_keys (bool) – don’t raise exception if key is missing.

  • kwargs – other arguments for the np.pad or torch.pad function. note that np.pad treats channel dimension as 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.

__init__(keys, spatial_border, mode=NumpyPadMode.CONSTANT, allow_missing_keys=False, **kwargs)[source]
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[Seq