Change history

The change history lists the most important changes and is not an exhaustive list.

0.3.4 (2024-11-22)

• Fix dependency on numpy v1 by using np.Inf instead of np.inf #52

0.3.3 (2024-09-26)

• Fix Jaccard metric computation on empty masks #49

• Fix the read the docs build

0.3.2 (2022-02-25)

• Minor bugfixes and usability improvements

• U-Net training examples in PyTorch and TensorFlow, see examples

• Improved documentation

0.3.1 (2020-08-02)

• Fixed missing dependency in setup.py

0.3.0 (2020-07-14)

• pymia.data package now supports PyTorch and TensorFlow. A few classes have been renamed and refactored.

• pymia.evaluation package with new evaluator and writer classes. Metrics are now categorized into pymia.evaluation.metric.categorical and pymia.evaluation.metric.continuous modules

• New metrics PeakSignalToNoiseRatio and StructuralSimilarityIndexMeasure

• Removed config, deeplearning, and plotting packages

• Improved readability of code

• Revised examples

• Revised documentation

Migration guide

Heavy changes have been made to move pymia towards a lightweight data handling and evaluation package for medical image analysis with deep learning. Therefore, this release is, unfortunately, not backward compatible. To facilitate transition to this and coming versions, we thoroughly revised the documentation and the examples.

0.2.4 (2020-05-22)

• Bug fixes in the pymia.evaluation package

0.2.3 (2019-12-13)

• Refactored: pymia.data.transformation

• Bug fixes and code maintenance

0.2.2 (2019-11-11)

• Removed the tensorflow, tensorboardX, and torch dependencies during installation

• Bug fixes and code maintenance

0.2.1 (2019-09-04)

• New statistics plotting module pymia.plotting.statistics (subject to heavy changes and possibly removal!)

• Bug fixes and code maintenance

• Several improvements to the documentation

0.2.0 (2019-04-12)

• New pymia.deeplearning package

• New extractor PadDataExtractor, which replaces the PadPatchDataExtractor (see migration guide below)

• New metrics NormalizedRootMeanSquaredError, SurfaceDiceOverlap, and SurfaceOverlap

• Faster and more generic implementation of HausdorffDistance

• New data augmentation module pymia.data.augmentation

• New filter BinaryThreshold

• Replaced the transformation in SubjectAssembler by a more flexible function (see migration guide below)

• Minor bug fixes and maintenance

• Several improvements to the documentation

We kindly appreciate the help of our contributors:

• Jan Riedo

• Yannick Soom

Migration guide

The extractor PadPatchDataExtractor has been replaced by the PadDataExtractor to facilitate the extraction flexibility. The PadDataExtractor works now with any kind of the three data extractors (DataExtractor, RandomDataExtractor, and SelectiveDataExtractor), which are passed as argument. Further, it is now possible to pass a function for the padding as argument to replace the default zero padding. Suppose you used the PadPatchDataExtractor like this:

import pymia.data.extraction as pymia_extr
pymia_extr.PadPatchDataExtractor(padding=(10, 10, 10), categories=('images',))

To have the same behaviour, replace it by:

import pymia.data.extraction as pymia_extr
pymia_extr.PadDataExtractor(padding=(10, 10, 10),
                            extractor=pymia_extr.DataExtractor(categories=('images',)))

The transformation in SubjectAssembler.add_batch() has been removed and replaced by the on_sample_fn parameter in the constructor. Replacing the transformation by this function should be straight forward by rewriting your transformation as function:

def on_sample_fn(params: dict):
  key = '__prediction'
  batch = params['batch']
  idx = params['batch_idx']

  data = params[key]
  index_expr = batch['index_expr'][idx]

  # manipulate data and index_expr according to your needs

  return data, index_expr

0.1.1 (2018-08-04)

• Improves the documentation

• Mocks the torch dependency to build the docs

0.1.0 (2018-08-03)

• Initial release on PyPI