Version 0.8.2

Changes w.r.t. Version 0.7.1:

Baseline algorithms:

• Crop very large cases for nnU-Net and nnDetection training
• Allow skipping validation inference
• UNet training flexibility
• UNet training Docker
• Combine labels in workdir
• Flexibility splits specification
• Bugfix additional training arguments

Evaluation:

• Allow evaluation without explicit subject list
• nnU-Net evaluation flexibility

Sagemaker:

• Restructure training on SageMaker: (#32)
  • Split preprocessing and training jobs
  • Add bash scripts to test preprocessing and training locally
• SageMaker training pipeline for semi-supervised nnU-Net
• Additional documentation SageMaker
• SageMaker training update
• Preprocessing cleanup

Preprocessing:

• Preprocessing flexibility (#34):
  • Convert preprocessing scripts (for supervised and semi-supervised learning) to importable functions.
  • Allow preprocessing of the data using the command line interface.
• Allow any split when resampling annotations (#37)

Documentation:

• Update Installation documentation

Bugfixes:

• Fix mlxtend version
• Fix version of dependencies

Version 0.7.1

• SageMaker training - part 1 (#18)

  • Training scripts for SSL nnU-Net on SageMaker
    • Improve the robustness of train.py to make debugging easier
    • Install dependencies of nnU-Net in SageMaker's requirements.txt
    • Install modified nnU-Net without requirements, to prevent requiring internet access
  • Convert U-Net plan_overview to a function
  • Training scripts for SSL U-Net on SageMaker
  • Add cross-validation splits with 10 cases for debugging
  • Add missing U-Net dependency

• Sagemaker training part2 (#26)

  • Data preprocessing: Improved specification of preprocessing settings,
    and logging of unexpected parameters
  • Improved training in distributed environments (such as SageMaker)
  • nnU-Net: improved preprocessing performance with nnUNet_tl
  • nnU-Net: preprocess scans to a maximum physical size of 81 x 192 x 192
    mm
  • U-Net: improve robustness of preprocessing, by skipping cases with
    label interpolation error
  • U-Net: add missing dependency
  • Cross-validation splits: add PI-CAI PubPrivTrain, unit-tests, and
    bugfix when using multiple at once

• Performance optimization

• PI-CAI PubPrivTrain cross-validation splits (#21)

• Improved logging preprocessing script

• Improve input specification preprocessing scripts

  • Location of the dataset and working direction can now be specified in three ways:
    • Using command line argument (i.e., --workdir=...)
    • Using environment variable (i.e. ENV workdir=..., for example in Dockerfile/docker run command/Python)
    • Mount folders to the default location (i.e., with -v flag in docker run)
  • Images and labels can now be located at separate locations.

• Update documentation

• Fix for nnU-Net inference softmax (#12)

  • Fix how nnUNet softmax predictions are converted from their .npz format of the cropped image to the physical extent of the original image
  • Evaluate while cropping away exterior predictions
  • Single configurable nnU-Net evaluation script

• Bugfixes

  • Fix UNet path handling (#19)
  • Update to correct FL gamma (149cd4d)

• Cleanup

  • Encapsulate plan_overview in a function
  • Command line options for preprocessing settings (#16)

Version 0.1

Initial release.