Abstract
Domain shift caused by heterogeneous scanners and imaging protocols remains a major obstacle to deploying medical image segmentation models across clinical centres. Source-free domain adaptation (SFDA) addresses this challenge by adapting a source-trained model to unlabeled target samples without accessing source data, thus complying with privacy constraints. While recent SFDA methods rely primarily on pseudo-label refinement or entropy minimization, directly optimizing large-scale foundation models under noisy self-training signals often leads to limited generalization and unstable convergence. In this work, we propose a meta-spatial adaptation framework built upon the Segment Anything Model (SAM) for prostate MRI segmentation, termed M-SFDA. Instead of updating the entire model, we introduce meta networks into the SAM’s image encoder to enable input-adaptive feature refinement. These meta networks explicitly model bidirectional spatial interactions along height and width axes, compensating for the limited spatial inductive bias of patch-based transformers. To further stabilize adaptation under unreliable pseudo supervision, we incorporate sharpness-aware optimization to encourage flatter minima and improve generalization. Extensive experiments on a six-domain prostate MRI benchmark demonstrate that M-SFDA consistently outperforms state-of-the-art SFDA approaches and achieves superior average Dice scores across diverse cross-domain scenarios. Our results highlight the potential of spatially adaptive strategies for medical foundation models under source-free settings.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MLMI_006.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=RkNDvQEkAj
BibTex
@InProceedings{FuYih_MSFDA_MICCAISAT2026,
author = { Fu, Yihang AND Chen, Ziyang AND Ma, Rongze AND Xia, Yong},
title = { { M-SFDA: Meta Network with Spatial Interaction for Source-Free Domain Adaptation in Prostate MRI Segmentation using Segment Anything Model } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17272},
month = {pending},
page = {pending}
}