ReLiSS

Inference, evaluation, completed-run logs and fixed data splits: hekaiwang/ReLiSS.

Training code and the complete experiment pipeline are not included in this release.

Dataset	Protocol	Checkpoints	Channel order
BraTS2021	Five folds, 1,251 subjects	fold 0–4, best	T1, T1ce, T2, FLAIR
ISLES2022	Five folds, 250 subjects	fold 0–4, final	FLAIR, ADC, DWI
WMH2017	Official 60 train / 110 test	fold all, final	FLAIR, T1

This repository contains eleven training checkpoints, matching plans, dataset metadata and splits. WMH uses the official train/test protocol; the training split file does not imply five released WMH models.

weights_manifest.json and SHA256SUMS record checkpoint sizes and SHA-256. The checkpoint tensors and training state are preserved. The trainer folder name is nnUNetTrainerReLiSS_300epochs; the supplied inference script directly constructs the matching network and loads its state dictionary.

Download

From the installed code repository:

source scripts/env.sh
python scripts/download_weights.py --repo wanghekai/ReLiSS \
  --revision COMMIT_HASH --output "$nnUNet_results"

Replace COMMIT_HASH with the immutable model revision in the code repository's docs/WEIGHTS.md. The downloader fetches exactly eleven weights and verifies sizes and hashes. Follow the code README for held-out inference and evaluation. For cross-validation, use each subject's held-out fold. Missing channels are zeroed after full-input preprocessing; the released protocol uses a shared crop.

Project code and model release are MIT licensed. Dataset and dependency terms remain separate. Obtain MRI data and reference masks from their authorized sources.