A Weakly Supervised Deep Learning Strategy for Assisting Malignancy Diagnosis in Urinary Cytology
Purpose Urinary cytology is a primary non-invasive screening tool for bladder cancer, but its diagnostic sensitivity remains suboptimal with significant inter-observer variability. This study aimed to develop and validate a weakly supervised deep learning model using multiple instance learning (MIL) for detecting malignancy in urinary cytology whole-slide images (WSIs), eliminating the need for labor-intensive cell-level annotations.
Methods A total of 3,042 urinary cytology WSIs were collected: 2,135 for training (1,884 benign, 251 malignant) from Uijeongbu St. Mary's Hospital and 907 for testing (600 benign from the same institution, 307 malignant from the Quality Assurance Program of the Korean Society of Cytopathology). The Clustering-constrained Attention MIL (CLAM) framework with a ResNet50 feature extractor was trained using slide-level labels only. Performance was evaluated using AUROC, accuracy, sensitivity, specificity, precision, and F1-score with 95% confidence intervals from bootstrap resampling.
Results The MIL model achieved an AUROC of 0.941 (95% CI, 0.924-0.957), accuracy of 0.860 (95% CI, 0.833-0.888), sensitivity of 0.824 (95% CI, 0.779-0.868), and specificity of 0.901 (95% CI, 0.862-0.937) on the external test dataset. Attention heatmap visualization confirmed that the model focused on cytologically suspicious regions corresponding to malignant cell clusters, providing interpretable diagnostic rationale.
Conclusions The weakly supervised MIL model demonstrated excellent diagnostic performance for malignancy detection in urinary cytology WSIs without requiring cell-level annotations. This approach offers a practical and scalable framework for AI-assisted cytopathology diagnosis, potentially improving screening efficiency and diagnostic reproducibility in clinical practice.