Abstract
1 Endoscopic ulcer assessment in Crohn’s disease (CD) is critical for clinical trials, yet manual annotation of potentially hour-long videos is impractical. Developing scalable ulcer segmentation models requires navigating a rapidly evolving design space spanning foundation model (FM) initialization, decoder architecture, and sequence-based models such as SAM. We present a large-scale empirical study of CD ulcer segmentation that evaluates four clinically relevant questions: (i) the impact of FM initialization on segmentation performance and label efficiency, (ii) generalization to unseen clinical trials, (iii) robustness across ulcer sizes, and (iv) whether SAM-based propagation can replace frame-wise segmentation in long endoscopy videos. We benchmark SAM3, MedSAM2, EndoFM, and our high-resolution DINOv3-based encoder pretrained on approximately 61M unlabeled endoscopy frames (EndoD3-HR) across Linear Probe, SegFormer, TransUNet, and EMCAD decoders. Our results show that (1) FM initialization strongly influences segmentation accuracy and label efficiency, with domain-specific pretraining providing the largest gains and high-resolution adaptation providing additional improvements; (2) models trained on multiple trials transfer effectively to an unseen trial without trial-specific adaptation; (3) performance varies across ulcer sizes, with strong results on medium and large ulcers that are most relevant for disease-burden assessment; and (4) although fine-tuned SAM-based models segment initialized frames well, propagation degrades over time due to endoscopy-specific challenges, making frame-wise segmentation the more robust solution for large-scale video analysis. Together, these findings provide practical guidance for FM selection and deployment of ulcer segmentation systems in CD clinical trial workflows.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MLMI_076.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=kkItmwElyq
BibTex
@InProceedings{GunFab_Optimizing_MICCAISAT2026,
author = { Gunderson, Fabio AND Chaitanya, Krishna AND Hirsch, Lukas AND Mobadersany, Pooya AND Parmar, Chaitanya AND Surace, Lindsey AND Yarandi, Shadi AND Ghanem, Louis R. AND Mansi, Tommaso AND Cula, Gabriela Oana AND Standish, Kristopher A. AND Damasceno, Pablo F.},
title = { { Optimizing Design Choices for Endoscopic Ulcer Segmentation: Foundation Models, Decoders, and SAM-based propagation } },
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}
}