Abstract
Accurate segmentation of ischemic stroke lesions from T1-weighted magnetic resonance imaging is challenging because infarcts vary substantially in size, morphology, intensity, and disease stage. Variations in scanners and acquisition protocols further limit generalization across imaging centers. To address the above-mentioned challenges, we developed Ladle-ResEncL, an effective three-dimensional deep learning framework for ischemic stroke lesion segmentation in T1-weighted MRI. The proposed method integrates a ResEnc-L-derived residual encoder with the encoder-dominant and abridged-decoder design principles of our recent Deep Ladle-Net framework, which outperformed five state-of-the-art methods and ranked third in the 2024 Universal Lesion Segmentation Challenge (ULS23), for fast universal 3D lesion segmentation on the chest–abdomen–pelvis CT scans to examine 10 types of lesions at once. The network contains seven resolution stages, a deep residual encoder, and an abridged decoder with one convolutional block per stage. To address the highly imbalanced lesion-size distribution, we introduce a lesion-size-weighted Dice and cross-entropy objective. Three-dimensional connected lesion components are identified from each reference mask, and smaller lesions receive larger voxel weights during optimization. T1-weighted images are resampled to 1-mm isotropic spacing and normalized using foreground-based Z-score normalization. The proposed model was trained on 1,308 cases and evaluated on 145 held-out validation cases. Using the official ISLES’26 evaluation metrics, the model achieved a Dice score of 0.604 ± 0.292, an absolute volume difference of 4.232 ± 7.351 mL, a lesion-wise F1 score of 0.473 ± 0.315, and a PR-AUC of 0.678 ± 0.311.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/ISLES_027.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=HoTjfZhNvv
BibTex
@InProceedings{WanChi_LadleResEncL_MICCAISAT2026,
author = { Wang, Ching-Wei AND Wang, Ting-Yi},
title = { { Ladle-ResEncL with Lesion-Size-Weighted Learning 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}
}