LI Yu, XU Xiaodan, ZHU Jinzhou
Objective To establish an interpretable computer model based on clinical knowledge, compare it with the traditional blackbox model, and further improve the effect of the model in the Mayo endoscopic score for ulcerative colitis(UC). 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 the internal validation set. The pictures were divided into Group 0, Group 1, Group 2 and Group 3 according to the Mayo endoscopic scoring system to construct the black box model. Meanwhile, all the pictures were marked according to five characteristics: ulcer (present or absent), spontaneous bleeding (present or absent), erythema (absent, visible, obvious), vascular texture (normal, blurred, absent), and mucosal fragility (normal, mild, brittle), and a sub-feature model was established. After fusion, an interpretable model was constructed. In addition, external validation was conducted using endoscopic images from the Endoscopy Center of Changshu Hospital Affiliated to Soochow University. In the external validation set, indicators such as calculation accuracy, matthew correlation coefficient (MCC), and Kapa coefficient were used to compare the performance of the interpretable model with that of the black-box model, and the classification index results of two endoscopists with different seniority were also compared. Finally, the Grad-CAM method was used to highlight the regions of the model reasoning basis. Results The accuracies of the four interpretable models based on MobileNet, ResNet, Xception and EfficientNet in the external validation set were 0.765, 0.800, 0.830 and 0.885 respectively. All were superior to the corresponding traditional black-box models of 0.665, 0.705, 0.775, and 0.815. Among them, the interpretability model based on EfficientNet performed best and was also superior to both junior physicians (0.805) and senior physicians (0.870). Conclusion In the Mayo endoscopic grading diagnosis of ulcerative colitis under endoscopy, the interpretable model performs better than the traditional deep learning black box model. Interpretable models have good application value in the endoscopic diagnosis of UC in the future.