Abstract
Automated segmentation of ischemic stroke lesions from MRI is a critical step toward scalable clinical decision support, yet performance on small and multifocal lesions remains a persistent bottleneck for existing deep-learning approaches. In this work, we present a baseline nnU-Net-based solution for the ISLES’26 challenge and conduct a size-stratified analysis of its failure modes. Using a 5-fold cross-validation protocol on the ISLES’26 training set (1453 cases), our baseline 3D full-resolution nnU-Net achieves an overall Dice score of 0.6465 (HD95 = 19.92 mm), but performance degrades substantially on small lesions (Dice = 0.3988) relative to large lesions (Dice = 0.7384). We characterize this bottleneck through a size-stratified evaluation and a dataset-level analysis showing that 66\% of cases present with multiple, spatially distinct lesion components. We further report a preliminary architectural exploration using a residual-encoder U-Net variant, which improves large- and medium-lesion Dice but does not resolve the small-lesion deficit, suggesting that representational capacity alone is not the limiting factor. We conclude by outlining biologically motivated directions - including contralateral hemispheric symmetry features and vascular territory priors - as promising avenues for future work on this specific failure mode.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/ISLES_032.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=Opi5zRQjTl
BibTex
@InProceedings{PatMeg_Investigating_MICCAISAT2026,
author = { Patel, Megh AND Karia, Konark},
title = { { Investigating Small-Lesion Segmentation Failures in Acute Ischemic Stroke: An nnU-Net Baseline Analysis for ISLES'26 } },
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}
}