Abstract
Efforts to adapt foundation models to medical imaging have yielded capable general-purpose segmentation systems, with models such as Medical Segment Anything Model (MedSAM) demonstrating strong generalisation across anatomies and modalities. Yet their general-purpose design leaves room for improvement on anatomically complex structures, particularly intricate shapes, fine boundaries, or fracture fragments, which are underrepresented in training data yet where diagnostic accuracy is most critical. In this work, we propose a lightweight Mixture of Experts (MoE) refinement module that operates on top of existing segmentation outputs. A gating mechanism routes each input to a corresponding specialist expert network based on anatomical regime (e.g., large vs. small structures), which then refines the initial segmentation mask. We validate our approach on a fractured pelvis X-ray dataset, a setting known to challenge general-purpose segmentation models. Our results demonstrate that routing to specialized experts improves segmentation quality across difficult regimes, and that shape-aware generative refinement is a promising direction for anatomically precise medical image segmentation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/ShapeMI_046.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/ShapeMI_046_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=64REK28ZDA
BibTex
@InProceedings{CheDen_Mixture_MICCAISAT2026,
author = { Checiu, Denisa AND Grazia, Giacomo AND Ramautar, Arshana AND Karajeanes, Daua AND Patel, Sanvritti AND Çevik, Ece AND Simion, Ioana},
title = { { Mixture of Experts for Fine-Grained Shape Refinement in Medical Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17259},
month = {pending},
page = {pending}
}