MIC-DKFZ/medicaldetectiontoolkit

Built on PyTorch with custom CUDA kernels for 3D-optimized Non-Maximum Suppression and RoIAlign, the framework handles dynamic patching and weighted consolidation of overlapping predictions across patches and test-time augmentations. It integrates the MIC-DKFZ batchgenerators library for extensive data augmentation and supports training from both bounding box and pixel-wise annotations through on-the-fly connected component labeling. Evaluation metrics include COCO mean average precision computed at both object and patient levels, with 2D/3D visualization outputs.

1,349 stars. No commits in the last 6 months.