RESEARCH WORK
Accepted: 2025-03-11
Objective To build explainable models based on clinical knowledge and compare them with traditional black-box models to further improve the effectiveness of Mayo endoscopic score of ulcerative colitis. Methods A total of 2251 endoscopic images from the Endoscopy Center of the First Affiliated Hospital of Soochow University were collected as the training set and internal validation set. The black box model was constructed by categorizing the images into groups 0, 1, 2, and 3 according to the Mayo endoscopic score. At the same time, all images were labeled according to five features, including ulceration (presence or absence), spontaneous hemorrhage (presence or absence), erythema (without, visible, obvious), vascular texture (normal, fuzzy, disappearing), and mucosal friability (normal, mild, easily friable), to create a sub-featured model, which was fused to construct an interpretable model. In addition, endoscopic images from the Endoscopy Center of Changshu Hospital affiliated with Soochow University were used for external validation. In the external validation set, the performance of the explainable model was compared with the black-box model by calculating the accuracy, Matthew correlation coefficient (MCC), and Cohen's Kappa with two endoscopists of different years of experience. Finally, the Grad-CAM method was used to highlight the regions on which the model's reasoning was based. Results The accuracies of the four explainable models based on MobileNet, ResNet, Xception, and EfficientNet in the external validation set were 0.765, 0.800, 0.830, and 0.885, respectively, which were better than those of the corresponding traditional black-box models, 0.665, 0.705, 0.775, and 0.815, with the explainable model based on EfficientNet performing the best, outperforming low seniority physicians (0.805) and high seniority physicians (0.870). Conclusion In this study, four explainable computer vision models were constructed and externally validated by collecting endoscopic images of ulcerative colitis, and the optimal model is selected, suggesting that the explainable model performs better than the traditional deep-learning black-box model in the Mayo endoscopic score of ulcerative colitis. The explainable model has good application value in the future endoscopic diagnosis of ulcerative colitis.