Abstract
Promptable foundation models have shown promise for medical image segmentation, but their potential and limitations for acute ischemic stroke lesion segmentation remain insufficiently understood. In this study, we evaluate MedSAM for diffusion MRI stroke lesion segmentation in a multi-center clinical cohort by separating the effects of prompt quality, stroke-domain adaptation, and automatic prompting. We compare zero-shot and adapted MedSAM variants with stroke-specific baselines. Under an oracle localization setting, where ground-truth bounding boxes and lesion-containing slices were derived from the reference masks, zero-shot MedSAM achieves moderate performance (Dice: 0.719 ± 0.138), but remains below 3D nnU-Net (Dice: 0.779 ± 0.143) and is sensitive to prompt quality. Under the same oracle setting, stroke-domain adaptation improves MedSAM, with full fine-tuning and lightweight encoder adapters reaching Dice scores of 0.829 ± 0.089 and 0.826 ± 0.093, respectively. However, when prompts are generated automatically, performance drops substantially (Dice: 0.587 ± 0.262), suggesting that automatic lesion localization is a major challenge for fully automatic MedSAM-based segmentation. Anatomical prior-guided prompting partially mitigates this limitation (Dice: 0.701 ± 0.177). These results suggest that MedSAM can be effective for acute stroke lesion segmentation when accurate prompts and domain adaptation are available, but robust automatic prompting remains necessary before reliable fully automatic deployment.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MI4MedFM_016.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=Jm744WiR0q
BibTex
@InProceedings{LiuMin_Evaluating_MICCAISAT2026,
author = { Liu, Mingtian AND Lartizien, Carole AND Chiaravalloti, Leonardo AND Chen, Bailiang AND Lapergue, Bertrand AND Berthezène, Yves AND Mazighi, Mikael AND Cho, Tae-Hee AND Frindel, Carole},
title = { { Evaluating the potential and limitations of MedSAM for acute stroke lesion segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17261},
month = {pending},
page = {pending}
}