Abstract
Ulcerative colitis (UC) is commonly assessed using the Mayo endoscopic subscore (MES), which classifies disease severity into four ordered categories. Despite its clinical utility, MES grading is prone to considerable intra- and inter-observer variability. To overcome this limitation, we develope ACD(2)W-InceptionNeXt, a deep learning-based computer-aided diagnosis (CADx) system that predicts MES from endoscopic video segments using InceptionNeXt with a progressive ACD(2)W-Loss. This design explicitly leveraged the ordinal nature of MES and enhanced discrimination between adjacent severity levels, a persistent challenge in ordinal regression. Prior to training and inference, frames were extracted from video segments and quality control to ensure the retention of informative frames. The system then operated in two stages: a frame-level predictor determined MES for individual frames, and a segment-level predictor subsequently aggregated frame-level outputs with predefined criteria to generate a final MES for each video segment. Experimental results demonstrated the performance across two UC endoscopic datasets. On the Changhua Christian Hospital (CCH) dataset, the system achieved 0.770 accuracy and a quadratic weighted kappa (QWK) of 0.873 at the frame-level, and 0.834 accuracy with QWK of 0.935 at the segment-level. On the Labeled Images for Ulcerative Colitis (LIMUC) dataset, it attained 0.797 accuracy and a QWK of 0.872 at the frame-level. These results showed that the proposed system could more effectively assist gastroenterologists in the estimation of UC severity.