ISLES'22 Stroke Lesion Segmentation — nnU-Net v2
Ischemic stroke lesion segmentation from DWI, ADC, and FLAIR MRI. Single-fold 3d_fullres nnU-Net v2, trained on ISLES 2022 (220 train / 30 held-out cases).
Code: GitHub — isles22-stroke-segmentation
Research use only, not a clinical or diagnostic tool.
Results (30-case internal holdout)
| Metric | Mean | Median |
|---|---|---|
| Dice | 0.804 | 0.861 |
| IoU | 0.702 | 0.756 |
| Hausdorff95 (mm) | 6.64 | 2.68 |
Two failure cases (strokecase0046, strokecase0030) pull the mean down. See the GitHub repo for full per-case metrics. Single fold, reduced-epoch trainer, not evaluated on the official ISLES'22 test set.
Usage
nnUNetv2_predict -i /path/to/imagesTs -o /path/to/output -d 001 -c 3d_fullres -f 0 -chk checkpoint_best.pth -m <downloaded_model_dir>
Channels: 0 = DWI, 1 = ADC, 2 = FLAIR.
Citation
@article{hernandez2022isles, title = {ISLES 2022: A multi-center MRI stroke lesion segmentation dataset}, author = {Hernandez Petzsche, Moritz R. and de la Rosa, Ezequiel and others}, journal = {Scientific Data}, volume = {9}, pages = {762}, year = {2022}}
@article{isensee2021nnunet, title = {nnU-Net: a self-configuring method for biomedical image segmentation}, author = {Isensee, Fabian and Jaeger, Paul F. and others}, journal = {Nature Methods}, volume = {18}, pages = {203--211}, year = {2021}}
Model by Mehmet Akif Karadağ. MIT License (code). Dataset: CC BY 4.0.