Interpretable machine learning for urothelial cells classification and risk scoring in urine cytology

Urine cytology is widely used for detecting urothelial carcinoma (UC), though its performance is constrained by limited sensitivity and substantial interobserver variability. An interpretable machine learning framework was developed to classify urothelial cells and to estimate slide-level risk of high-grade UC. 10,230 expert-annotated urothelial cells were used to extract 20 quantitative feature representing cytomorphologic criteria defined by the Paris System. Ordinal logistic regression and random forest models were trained and validated, achieving over 90% accuracy for classifying cells into normal, atypical, or suspicious categories. Interpretable morphological features were identified as major contributors to prediction. Slide-level risk scores were derived from aggregated cell probabilities in a validation set of 247 cases. These scores effectively stratified negative, atypical, low-grade, and high-grade UC cases (p < 0.0001). Through alignment with established cytologic criteria, this feature-based framework provides a transparent and quantitative approach that may improve consistency, efficiency, and interpretability in digital urinary cytology.