wmh-mc-seg
Automatic segmentation of white matter hyperintensities from brain MRI with uncertainty estimation. Published in Computers in Biology and Medicine (2025).
Segments white matter hyperintensities (WMH) from T1 + FLAIR brain MRI and produces per-voxel uncertainty maps. The core question: can entropy-based regularization make the model’s uncertainty a reliable proxy for segmentation errors, especially under domain shift? Trained and evaluated on the WMH Segmentation Challenge dataset (Utrecht, Amsterdam, Singapore).
Publication
Matzkin, F. et al. “Towards reliable WMH segmentation under domain shift: An application study using maximum entropy regularization to improve uncertainty estimation.” Computers in Biology and Medicine, vol. 196, Part A, 110639, 2025. DOI
The published code corresponds to the v0.1.0 release. The current version modernizes the codebase (Lightning 2.x, PyTorch 2.x, MONAI, proper packaging, tests, CI) without changing the model architecture or training logic.
What it does
- Segmentation: Patch-based 3D U-Net trained with configurable loss functions (CE, Dice) and uncertainty regularizers (MEEP, KL, MEALL)
- Uncertainty estimation: Monte Carlo dropout at inference produces per-voxel uncertainty maps alongside the segmentation
- Analysis: Marimo apps for interactive slice viewing and metric dashboards (Dice vs entropy, calibration, per-center breakdowns)
Technologies
- Core: Python, PyTorch, Lightning, MONAI, TorchIO
- Analysis: Marimo, SimpleITK, NiBabel
- Experiment tracking: MLflow