v0.4.20-alpha

Bug fixes for VAE regularization.

Bug fixes for pre-trained models. We know this works for nnUNet models, but we need to do more investigation.

Refactor analysis, preprocessing, inference, and training modules to production-quality code.

v0.4.19-alpha

Bug fix for evaluation.

Readability for util functions related to creating paths CSV files.

Filter out hidden files when creating train_paths.csv and fg_bboxes.csv.

Add extra error handling for small datasets and multiple GPUs.

v0.4.18-alpha

Refactor the analyze_data module to have production-quality code.

Change the back_to_original_space function in main_inference to set the predictions spacing to the target spacing before applying resampling to the original spacing.

v0.4.16-alpha

Add cldice loss function and include dynamic weighting with dice+ce loss function.

Change validation loss from dice to dice+ce.

v0.4.15-alpha

Update calculation of steps per epoch.

v0.4.12-alpha

Update the training loop so that testing on each fold happens after training. This addresses a known bug that occurs due to the GPU pool sitting idle for too long while the rank 0 GPU runs inference on the test set for a given fold.

Improvements to the computation of the DTMs to handle cases where a current class does not exist in a mask and an option to normalize the DTMs so that their values lie between -1 and 1.

v0.4.11-alpha

The bug fix for inference failed. Reverting to the previous version of the back_to_original_space function in main_inference.py.

v0.4.10-alpha

Update license to Apache 2.0.

v0.4.9-alpha

Update inference for a bug in ants.reorient_image2

v0.4.8-alpha

Upgrade to PyTorch 2.3 and add flag to set surface dice tolerance