Quantitative image feature integrated digital pathology model based on deep learning and machine learning for binary histological severity stratification of ulcerative colitis

The STRIDE II consensus recognizes histopathological healing in ulcerative colitis as a key therapeutic target. This study aimed to develop and internally evaluate a quantitative image-based model for the binary classification of mild versus moderate-to-severe UC based on histopathological features. Pathological slides from 167 ulcerative colitis patients (58 mild, 109 moderate-to-severe) were retrospectively collected from two medical centers. A U-Net model quantified glandular structures (number, area ratio), Gray-Level Co-occurrence Matrix analyzed texture features, and a five-interval grayscale histogram was modeled. Features selection was performed using LASSO regression. Patients were randomly divided into training and test sets in a 7:3 ratio. Machine learning models were constructed and their performance was evaluated. In the five-fold cross-validation analysis within the training set, SVM achieved the highest mean ROC-AUC and was selected as the final model. After refitting on the full training set, the final SVM model achieved an AUC of 0.891 (95% CI: 0.742–0.958), a sensitivity of 0.848 (95% CI: 0.691–0.933), a specificity of 0.833 (95% CI: 0.608–0.942), and an F1-score of 0.875 (95% CI: 0.774–0.952) on the independent test set. SHAP analysis identified gland area ratio, fibrosis, and GrayHist_100_149, defined as the proportion of mucosal pixels with intermediate grayscale intensity ranging from 100 to 149 in grayscale-converted H&E images, as the most important predictors. This study integrated glandular geometry, texture features, and grayscale analysis using machine learning to develop a quantitative auxiliary tool for the binary stratification of mild versus moderate-to-severe UC. These findings suggest that this approach may serve as a complementary approach for histological assessment in ulcerative colitis.