Abstract
Pipelines for chronic ischaemic stroke lesion segmentation often accumulate normalisation, sampling, architectural, and post-processing components whose individual contributions remain unclear. We examine these choices through staged development on multicentre T1-weighted magnetic resonance imaging (MRI), yielding a 3D DynUNet pipeline with raw intensities, $128^3$-voxel patches, spatial augmentation, and Dice–focal loss, without atlas registration or threshold calibration. On the ATLAS v2.1 development split, leading DynUNet configurations attain Dice scores of 0.625–0.633, compared with 0.601 for an equivalently evaluated nnU-Net baseline. However, on ATLAS v3.0 three-fold cross-validation, nnU-Net attains a higher mean Dice than DynUNet (0.628 vs 0.610), reversing the development ranking; DynUNet records a slightly lower mean 95th-percentile Hausdorff distance. The three fold-specific DynUNets form the ISLES 2026 submission ensemble, whose hidden-test performance remains unavailable. Normalisation, aggressive lesion oversampling, and increased capacity provide no consistent development benefit. These findings show that promising pipeline choices may not transfer across dataset expansions, underscoring the importance of strong baselines and validation beyond a single development split.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/ISLES_051.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=oz9jvOtIcC
BibTex
@InProceedings{NesMin_Evaluating_MICCAISAT2026,
author = { Nessiem, Mina A. AND Schuller, Björn W.},
title = { { Evaluating Pipeline Design Choices for Chronic Stroke Lesion Segmentation in T1-Weighted MRI } },
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}
}
