What’s new in 1.2#
Auto3DSeg enhancements and benchmarks
nnUNet integration
TensorRT-optimized networks
MetricsReloaded integration
Bundle workflow APIs
Modular patch inference
Lazy resampling for preprocessing
Auto3DSeg enhancements and benchmarks#
Auto3DSeg is an innovative solution for 3D medical image segmentation, leveraging the advancements in MONAI and GPUs for algorithm development and deployment. Key improvements in this release include:
Several new modules to the training pipelines, such as automated GPU-based hyperparameter scaling, early stopping mechanisms, and dynamic validation frequency.
Multi-GPU parallelism has been activated for all GPU-related components including data analysis, model training, and model ensemble, to augment overall performance and capabilities.
The algorithms were benchmarked for computational efficiency on the TotalSegmentator dataset, containing over 1,000 CT images.
Multi-node training is implemented, reducing model training time significantly.
nnUNet integration#
The integration introduces a new class, nnUNetV2Runner
, which leverages Python APIs to facilitate model training, validation,
and ensemble, thereby simplifying the data conversion process for users.
Benchmarking results from various public datasets confirm that nnUNetV2Runner performs as expected.
Users are required to prepare a data list and create an input.yaml
file to install and use the system.
The framework also allows automatic execution of the entire nnU-Net pipeline, from model training to ensemble,
with options to specify the number of epochs. Users can access APIs for training, dataset conversion, data preprocessing, and other components.
Please check out the tutorials for more details.
TensorRT-optimized networks#
NVIDIA TensorRT is an SDK for high-performance deep learning inference,
includes a deep learning inference optimizer and runtime that delivers low latency and high throughput for inference applications.
It can accelerate the deep learning model forward computation on the NVIDIA GPU.
In this release, the trt_export
API to export the TensorRT engine-based TorchScript model has been integrated into the MONAI bundle.
Users can try to export bundles with it. A few bundles in the MONAI model zoo,
like the spleen_ct_segmentation
and endoscopic_tool_segmentation bundles,
have already been exported and benchmarked. For more details about how to export and benchmark a model,
please go to this tutorial.
MetricsReloaded integration#
MetricsReloaded - a new recommendation framework for biomedical image analysis validation - is released publicly via https://github.com/Project-MONAI/MetricsReloaded. Binary and categorical metrics computing modules are included in this release, using MetricsReloaded as the backend. Example scripts are made available to demonstrate the usage.
Bundle workflow APIs#
BundleWorkflow
abstracts the typical workflows (such as training, evaluation, and inference) of a bundle with three main interfaces:
initialize
, run
, and finalize
, applications use these APIs to execute a bundle.
It unifies the required properties and optional properties for the workflows, downstream applications
can invoke the components instead of parsing configs with keys.
In this release, ConfigWorkflow
class is also created for JSON and YAML config-based bundle workflows for improved Pythonic usability.
Modular patch inference#
In patch inference, patches are extracted from the image, the inference is run on those patches, and outputs are merged to construct the result image corresponding to the input image. Although depending on the task, model, and computational/memory resources, the exact implementations of a patch inference may vary, the overall process of splitting, running inference, and merging the results remains the same. In this release, we have created a modular design for patch inference, which defines the overall process while abstracting away the specific behavior of how to split the image into patches, how to pre and post process each patch, and how to merge the output patches.
Lazy Resampling for preprocessing#
Lazy Resampling is a new, experimental feature for preprocessing. It works under the hood along with MONAI transforms to combine adjacent spatial and cropping transforms into a single operation. This allows MONAI to reduce the number of data resamples a pipeline undergoes. Depending on the preprocessing pipeline, it can potentially:
reduce processing time
reduce processing memory
reduce incidental artifacts added by resampling
preserve data that would otherwise be cropped and replaced with padding
Lazy Resampling pipelines can use a mixture of MONAI and non-MONAI transforms, so
should work with almost all existing pipelines simply by setting lazy=True
on MONAI Compose
instances. See the
Lazy Resampling topic
in the documentation for more details.