Abstract
Colonoscopy is central to colorectal cancer prevention, yet building robust AI systems for detection, segmentation, and diagnosis remains constrained by the cost of expert annotation and the scale of video data. In colonoscopy, clinically relevant representations must capture both global semantics and fine-grained local detail: lesions are small and transient, yet must be recognized robustly under substantial appearance variation caused by illumination changes, specularities, and debris. Self-supervised learning (SSL) can exploit abundant unlabelled endoscopy video, but existing approaches only partially address these requirements — joint-embedding methods learn strong global semantic representations but under-emphasize local spatial detail, while masked modelling approaches encourage fine-grained spatial representations but can bias learning toward low-level statistics through pixel-level reconstruction targets. We propose STOMP (Semantic Token-dropping Masked Pre-training), a novel SSL framework designed for the unique demands of endoscopy video, which jointly optimises global and local semantic learning within a unified token-dropping encoder. Our method achieves the efficiency critical for video pre-training while simultaneously preserving the semantics required for clinically meaningful representations — without pixel reconstruction or doubled encoder passes. Results across classification, segmentation, and detection show that STOMP yields strong colonoscopy video representations while requiring substantially less pre-training compute compared to prior methods.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/CaPTion_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/profile?id=~Nils_Hallerfelt1
BibTex
@InProceedings{HalNil_Efficient_MICCAISAT2026,
author = { Hallerfelt, Nils L. AND Cherubini, Andrea AND Biffi, Carlo AND Tarroni, Giacomo},
title = { { Efficient Self-Supervised Pre-Training in Endoscopy } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17278},
month = {pending},
page = {pending}
}