What question did this study set out to answer?

To evaluate lightweight models for predicting the severity of ulcerative colitis using stool images.

March 6, 2026The Journal of the Korean Institute of Information and Communication Engineering0 citationsOpen Access

Ulcerative Colitis Severity Prediction Using Multiple Lightweight Models for On-Device Deployment

DWDongwon WooKyungpook National University Hospital BKBumhwi Kim HCHo-Young Chung

Puntos clave

To evaluate lightweight models for predicting the severity of ulcerative colitis using stool images.
Implemented lightweight models: MobileNetV2, EfficientNet-B0, ShuffleNetV2, GhostNet.
Compared model performance in relation to the Vision Transformer (ViT).
Restructured UCEIS scores into three classes for multi-class classification.
Lightweight models performed effectively in predicting UC severity from stool images.
MobileNetV2 and EfficientNet-B0 showed the best performance and efficiency.
The multi-class classification approach enhances clinical relevance.

Resumen

Ulcerative colitis (UC) severity is typically evaluated using the Ulcerative Colitis Endoscopic Index of Severity (UCEIS), but endoscopy is invasive and costly. To explore a non-invasive alternative, previous work proposed a Vision Transformer (ViT)-based model using stool images for binary classification of UCEIS scores. In this study, we extend this approach in two ways. First, to enable practical use in resource-limited and on-device settings, we implement lightweight models—MobileNetV2, EfficientNet-B0, ShuffleNetV2, and GhostNet—and compare their performance and efficiency against ViT. Second, we restructure UCEIS scores into three clinically relevant classes (0–1, 2–4, 5–8) for multi-class classification. This work demonstrates the feasibility and clinical value of stool image-based UC

Preguntar a la IA

Me gusta

Guardar

Ver artículo completo