Advanced AI Model for Predicting NMIBC Recurrence: Integrating Cystoscopy Images and EMR Data

PURPOSE: Non–muscle invasive bladder cancer (NMIBC) is prone to recurrence, requiring long-term surveillance and accurate risk prediction. While cystoscopy is essential in monitoring, conventional scoring systems (e.g., EORTC, CUETO, EAU) rely on limited clinical factors and may miss complex visual and clinical patterns. This study aimed to develop an AI model that integrates cystoscopic images and clinical data to improve recurrence prediction.
METERIALS AND METHOD: We retrospectively collected 2,256 cystoscopy images from 277 patients with pathologically confirmed NMIBC treated at Chung-Ang University Hospital (2006–2022). The cohort included 66 females and 211 males (median age, 70 years; IQR, 64–76). Clinical variables included demographics, smoking history, lab results, tumor characteristics, and recurrence outcomes. A multi-input model combined image features from a modified EfficientNetB0 and clinical features via a dense neural network. Thirty cases were used for evaluation. Model performance was assessed using AUROC, accuracy, sensitivity, specificity, and precision. Long-term metrics included time-dependent AUC and Integrated Brier Score (IBS). Results were compared with EORTC, CUETO, and EAU scores.
RESULTS: The Multi-Input NMIBC Recurrence (MIBR) model achieved an AUROC of 84.3%, outperforming EORTC (54.8%), CUETO (59.8%), and EAU (65.2%). At 5 and 10 years, AUCs were 81.8% and 98.0%, respectively. The model had the lowest IBS (0.1598), with all results statistically significant (p < 0.001).