Domain Adaptation for Ulcerative Colitis Severity Estimation Using Patient-Level Diagnoses

2025-09-17

AI Summary

Unsupervised/semi-supervised domain adaptation methods underperform in ulcerative colitis (UC) severity assessment due to domain shift arising from heterogeneous imaging equipment and clinical protocols across hospitals, especially when target-domain annotations are scarce. Method: We propose a weakly supervised domain adaptation framework leveraging only coarse-grained patient-level diagnostic labels (e.g., “mild,” “moderate,” “severe”). It employs a shared aggregation token mechanism to model global lesion distribution and introduces a maximum-severity triplet loss to enforce alignment of class-level feature distributions by focusing on the most discriminative, severely affected regions. Contribution/Results: Our method significantly reduces annotation burden while achieving an average 4.2% improvement in severity classification accuracy over state-of-the-art domain adaptation approaches on a multi-center UC endoscopic image dataset, demonstrating strong effectiveness and generalizability in realistic cross-domain settings.

Technology Category

• Machine Learning: Transfer, Domain Adaptation, Multi-Task Learning
• Computer Vision: Segmentation
• Application Domains: Security

Application Category

• Search and Retrieval-Augmented AI: Vertical and domain-specific search
• Web Mining and Content Analysis: Robustness and generalizability of Web mining methods
• Economics, Online Markets and Human Computation: Data quality aspects of human-annotated datasets

Abstract

The development of methods to estimate the severity of Ulcerative Colitis (UC) is of significant importance. However, these methods often suffer from domain shifts caused by differences in imaging devices and clinical settings across hospitals. Although several domain adaptation methods have been proposed to address domain shift, they still struggle with the lack of supervision in the target domain or the high cost of annotation. To overcome these challenges, we propose a novel Weakly Supervised Domain Adaptation method that leverages patient-level diagnostic results, which are routinely recorded in UC diagnosis, as weak supervision in the target domain. The proposed method aligns class-wise distributions across domains using Shared Aggregation Tokens and a Max-Severity Triplet Loss, which leverages the characteristic that patient-level diagnoses are determined by the most severe region within each patient. Experimental results demonstrate that our method outperforms comparative DA approaches, improving UC severity estimation in a domain-shifted setting.

Problem

• Addressing domain shift in ulcerative colitis severity estimation
• Leveraging patient-level diagnoses as weak supervision
• Improving cross-domain severity estimation without costly annotations

Innovation

• Weakly Supervised Domain Adaptation method
• Shared Aggregation Tokens alignment
• Max-Severity Triplet Loss utilization