Abstract

Automated segmentation of ischemic stroke lesions from structural MRI is challenging due to lesion variability and pronounced class imbalance. We present our contribution to the ISLES’26 challenge, systematically evaluating four nnU-Net-based configurations on 1453 T1-weighted cases. To isolate architectural and data-level effects, all models shared identical preprocessing (reorientation and N4 bias correction). The evaluated configurations, all implemented within the nnU-Net framework, include: (1) the standard baseline, (2) a residual encoder variant, (3) the residual encoder variant trained with extended data augmentation, and (4) the residual encoder variant trained with an explicit bilateral (left-right) asymmetry input channel. On our internal hold-out test split, the residual encoder provided the strongest foundational improvement in median overlap and boundary accuracy. Building on this, the extended augmentation configuration achieved the best overall boundary precision (lowest HD95) and reduced missed lesion voxels. Conversely, while the bilateral asymmetry channel yielded mixed results on the internal test set, it demonstrated superior generalization on external data, outperforming the standalone residual encoder on the preliminary ISLES’26 leaderboard.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/ISLES_028.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=Hs8smEehRx

BibTex

@InProceedings{ZayBey_From_MICCAISAT2026,
        author = { Zayim, Beyza AND Aissiou, Ikram AND Collins, Stephan AND Lalande, Alain AND Meriaudeau, Fabrice},
        title = { { From Baseline to Bilateral Asymmetry: A Systematic nnU-Net Comparison for Ischemic Stroke Lesion Segmentation in 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}
}


back to top