isles26-nnunet-d507-topk10

ISLES'26 stroke lesion segmentation - nnU-Net v2 (Dataset507, Dice+TopK10)

Inference-only weights for the ISLES'26 submission container. 5-fold 3d_fullres nnU-Net v2 trained on the ISLES'26 public release (ATLAS R3.0 derived, 1452 cases) with a center-grouped split: every fold holds out whole acquisition centers, which is the shift the hidden test set actually poses.

Contents

plans.json, dataset.json
fold_0..fold_4/checkpoint_final.pth
hf_manifest.json    # sha256 of every file; re-verified in CI and at image build

Checkpoints are stripped to network_weights, init_args, trainer_name and inference_allowed_mirroring_axes; optimizer state removed, tensor identity verified against the originals at export.

Measured performance

Out-of-fold on the center-held-out split (n=1452), scored with the ISLES'26 organizers' own evaluation code, at the shipped operating point (threshold 0.35, min connected component 20 voxels, 26-connectivity):

Caveats worth carrying: on the center-LEAKING random-fold split the same model scores Dice 0.6558, so 0.628 is the honest figure. The volume difference above is ~0.4 mL optimistic because it was measured on stored validation softmax rather than the deployment path.

Usage

Consumed by the ISLES'26 submission container, which handles native-space orientation round-tripping and emits both a binary mask and a float32 probability map.