Abstract
MedSAM offers a strong prior for medical image segmentation, but its transformer encoder and prompt-based inference make it difficult to deploy as an automatic model in constrained clinical settings. We propose TopoMedKD, a distillation framework that uses frozen, zero-shot MedSAM that is prompted with GT bounding box only during training and transfers its knowledge to compact, and automatic prompt-free student at inference time. The method combines three complementary modules: logit distillation, prototype-guided contrastive distillation, and a shape- and topology-aware structural loss based on signed-distance, soft-skeleton, and boundary consistency. A per-sample Focal-Tversky and Dice loss remains the dominant supervision term, limiting the risk of copying errors from the zero-shot teacher. On KiTS19 kidney tumour CT segmentation, TopoMedKD consistently improves lightweight students over supervised training, with the largest gain observed for ShuffleNetV2. The results suggest that a frozen foundation model can be useful for efficient medical segmentation when the transfer objective is designed to preserve both class structure and region shape.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MedAGI_023.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=UWb3WBFC9c
BibTex
@InProceedings{BenEdd_TopoMedKD_MICCAISAT2026,
author = { Ben Loussaief, Eddardaa AND Ayad, Mohammed AND Rashwan, Hatem A. AND Puig, Domenec},
title = { { TopoMedKD: Shape- and Topology-Aware Distillation of Zero-Shot MedSAM for Compact Medical Image 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}
}