Abstract
Ischemic stroke lesions in T1-weighted MRI span several orders of magnitude in size but mostly occupy only a fraction of the brain, so a single full-volume neural network spends most of its capacity on healthy tissue. We address this with a two-stage coarse-to-fine cascade built on nnU-Net: a first stage segments the lesion in the full volume, its prediction defines a tight bounding box, and a second stage refines the segmentation inside the cropped region of interest. Since ISLES’26 ships per-case metadata, we additionally condition the network on it through a Metadata-Informed Bottleneck, fusing days post-stroke, chronicity and a learned per-site embedding into the deepest encoder stage. A reserved unknown-site embedding, trained by dropping the site identity for 5\% of samples, keeps the model usable on unseen centers. In an ablation, metadata conditioning raises Dice from 60.8 to 62.0 and cropping adds a further large gain to 71.4. On the full training data, cross-validated Dice rises from 65.1 to 73.0 on the same cases and the 95th-percentile Hausdorff distance is halved from 18.5 to 9.3\,mm. Because the second stage inherits any localization failure of the first, we quantify this error propagation: the first-stage box retains 91.7\% of the lesion volume on average and loses it entirely in only 3.5\% of cases, all of which the single-stage model had already missed.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/ISLES_008.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/ISLES_008_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=kN1q2nZiu8
BibTex
@InProceedings{TapEli_AMetadataInformed_MICCAISAT2026,
author = { Tappeiner, Elias AND Gapp, Christian},
title = { { A Metadata-Informed Two-Stage nnU-Net Cascade for Ischemic 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}
}