MIST: Medical Imaging Segmentation Toolkit

Several medical image segmentation papers claim to beat nnU-Net, and at least one later study disputes those results. Part of the problem is that every group trains, preprocesses and evaluates differently. MIST (Medical Imaging Segmentation Toolkit) is an end-to-end framework that standardizes those pipelines so methods can be compared fairly on the same data.

What it does

• Data analysis: rule-based selection of foreground cropping, target spacing, patch size and normalization, saved to a config file.
• Preprocessing and training: the same pipeline for any architecture and loss function, runnable in one command and scalable across multiple GPUs.
• Evaluation: Dice, Hausdorff distance and surface Dice for any user-defined class, plus postprocessing and test-time inference.

The paper demonstrates it on the BraTS adult glioma post-treatment challenge data.

My part

Co-author, from my time at MD Anderson. I developed the clDice loss variants and handled training on the GPU cluster. I then applied MIST to biliary duct segmentation as first author of an RSNA abstract, written up separately.