Abstract
Colon capsule endoscopy (CCE) enables minimally invasive colorectal polyp detection, but class imbalance and patient-specific variations in image quality make accurate polyp characterisation, i.e., distinguishing neoplastic from non-neoplastic lesions, challenging. Selfsupervised learning with parameter-efficient fine-tuning provides strong generalisation but does not address the distributional gap introduced by class imbalance, leaving sensitivity to neoplastic lesions insufficient for clinical deployment. Classical augmentation fails to close this gap, as geometric and colour transforms do not introduce new visual patterns for the underrepresented class. To address this, we investigate Stable Diffusion-based synthetic image generation as a targeted augmentation strategy, evaluated against classical augmentation and no-augmentation baselines. Generated images are verified against a pre-specified StructuralSimilarityIndexMeasureacceptancecriteriontoconfirmthatevery synthetic sample preserves clinically relevant structure while introducing sufficient diversity. Two transformer-based models, DINOv3 and ViT- B/16,arefine-tunedusinglow-rankadaptationonaCCEdatasetcollected using the PillCam Colon 2 system. Stable Diffusion augmentation substantially improves DINOv3+LoRA characterisation performance, with sensitivity increasing from 84.74% to 96.32% and accuracy from 88.97% to 94.72%, where classical augmentation produced no measurable gain. These findings demonstrate that generative augmentation targeting class imbalanceisnecessarytoachieveclinicallyacceptablesensitivityforpolyp characterisation in CCE.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/SASHIMI_064.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=aTge2PvDt5
BibTex
@InProceedings{MouMoh_Synthetic_MICCAISAT2026,
author = { Moustapha, Mohamed-Wali Ahmed AND Mustafa, Florentin AND Bathineni, Ashajyothi AND Gogineni, Vinay Chakravarthi AND Koulaouzidis, Anastasios AND Nadimi, Esmaeil S.},
title = { { Synthetic Data Augmentation via Stable Diffusion for Polyp Characterisation in Colon Capsule Endoscopy } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17258},
month = {pending},
page = {pending}
}