v0.1.7-beta

Minor performance improvements for loss computation.

Add the learning rate to the training progress bar.

Add the option to do validation periodically after a certain number of epochs with the --validate-every-n-epochs and --validate-after-n-epochs flags.

v0.1.5-beta

Add MedNexT models to MIST. These models are also compatible with the --pocket flag.

Minor improvements to the dataloader for multiple classes.

Minor improvements and improved error handling for the Analyzer class.

v0.1.4-beta

Update the data loader to use class weights to select the foreground bounding box. This should help with datasets with a lot of labels.

v0.1.3-beta

Change analyze behavior to warn users of examples exceeding a maximum recommended memory size instead of coarsening the target spacing if such images are detected.

v0.1.2-beta

Bug fix for the case of bad data in the analysis pipeline. We needed to reset the index in the paths data frame.

v0.1.1-beta

This release includes major changes:

1. We updated the default number of epochs to 250000/steps_per_epoch. This allows MIST to run for 250,000 optimization steps. We also validate every 250 on the entire held-out fold for cross-validation. This makes comparisons to nnUNet easier.

2. Remove the generalized dice loss (GDL), the GDL with cross-entropy loss, and remove class weights from the generalized surface loss. We have not seen any advantage to using class weights in loss functions.

3. Update documentation to include the clDice.

4. Start adding the MIST version number in the config file.

v0.1.0-beta

This release includes major changes:

1. We updated the default number of epochs to 250000/steps_per_epoch. This allows MIST to run for 250,000 optimization steps. We also validate every 250 on the entire held-out fold for cross-validation. This makes comparisons to nnUNet easier.

2. Remove the generalized dice loss (GDL), the GDL with cross-entropy loss, and remove class weights from the generalized surface loss. We have not seen any advantage to using class weights in loss functions.

3. Update documentation to include the clDice.

4. Start adding the MIST version number in the config file.

v0.0.3-beta

Update the version to get the build badge to show passing. Probably not the best way to do this, but it'll do for now. This is the same code as v0.0.2-beta.

v0.0.2-beta

Improved error handling for preprocessing and conversion tools.

Improved computation of target spacing.

v0.0.1-beta

Fix orientation issue for inference.

This was the last major bug that we needed to fix before moving from alpha to beta versions of MIST.