Abstract
Manual tracing of stroke lesions is slow and varies between raters, limiting how many scans a study can cover. Diffusion-weighted MRI defines acute infarction most clearly but is rarely available later after stroke, whereas T1-weighted MRI is commonly acquired throughout follow-up. ISLES’26 ranks submissions on five metrics within each scan and averages those ranks, so Dice alone does not represent the full evaluation criterion. We therefore used the complete criterion to select training length, loss function, network configuration, and post-processing. We evaluated the resulting self-configuring nnU-Net by five-fold cross-validation across 1,453 subjects, stratified by centre, lesion size, and time since stroke. The model achieved a median Dice of 0.77, lesion-wise F1 of 0.67, and PR-AUC of 0.88. Selection under the complete criterion changed the preferred configuration: a hard-example loss ranked higher despite a lower median Dice than the alternative, and the criterion retained a small-component filter that Dice alone would have rejected, although selecting and evaluating the filter on the same fold inflated its apparent benefit. Error analysis showed that the evaluation rule can assign the same zero score to different failure modes. Of 148 scans scoring zero, 113 contained a predicted lesion, including 50 with predictions that overlapped the reference but did not reach the IoU matching threshold of 0.25. Pooled lesion-wise recall was lower than precision (0.41 vs. 0.64), indicating that missed lesions were the dominant error. Together, these results show that model selection based on the complete evaluation criterion can lead to different choices than Dice alone, while separate analysis of the underlying metrics identifies failure modes that aggregate scores can obscure.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/ISLES_039.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=ITkIVayPW9
BibTex
@InProceedings{GurOre_Rankingaware_MICCAISAT2026,
author = { Gurevitch, Oren AND Mitsis, Georgios D.},
title = { { Ranking-aware model selection for stroke lesion segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17255},
month = {pending},
page = {pending}
}