Abstract

Automatic polyp segmentation is crucial for improving the clinical identification of colorectal cancer. While pretrained representations are reshaping computer vision, their role in medical image segmentation under domain shift remains underexplored. We propose a representation-driven framework that leverages self-attention “key” features from DINO-pretrained Vision Transformers, which capture object-centric and spatially consistent representations, enabling more robust segmentation without relying on complex task-specific architectures. We evaluate our framework on a multi-center dataset with substantial inter-center variability, using Domain Generalization (DG) and Extreme Single Domain Generalization (ESDG) protocols to simulate realistic distribution shifts. Frozen DINO features achieve performance comparable to specialized architectures, while supervised fine-tuning consistently improves results, reaching state-of-the-art performance particularly in data-scarce settings. A systematic comparison of DINO variants (V1 to V3) further shows that advances in self-supervised representation learning translate into improved cross-domain generalization. These findings suggest that representation quality, rather than architectural complexity, is the primary driver of generalization in medical image segmentation. Code: https://github.com/Trustworthy-AI-UU-NKI/RepMatters.git.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MLMI_008.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=U1Qwt2UWlm

BibTex

@InProceedings{MonCar_Representation_MICCAISAT2026,
        author = { Monteiro, Carla AND Corbetta, Valentina AND Beets-Tan, Regina AND Teixeira, Luís F. AND Silva, Wilson},
        title = { { Representation Matters: Rethinking Domain Generalization in Polyp Segmentation } },
        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}
}


back to top