Explainable multimodal deep learning for recurrence prediction in BCG-treated non-muscle-invasive bladder cancer: a retrospective cohort study

Lee, Ju Young BS^a,b; Lee, Yong Seong MD, PhD^c; Lee, Jae Hyeok BS^d; Jung, Gu Cheol BS^d;Choi, Se Young MD, PhD^a,e,*

Background: 

Non-muscle-invasive bladder cancer (NMIBC) has a high risk of recurrence, but existing tools such as EORTC, CUETO, and EAU have a limited ability to assess response to standard intravesical Bacillus Calmette–Guérin (BCG). We aimed to develop and validate a multimodal deep learning model for predicting recurrence and timing in this BCG-treated population and to compare its performance with established clinical scoring systems.

Methods: 

 

We retrospectively analyzed 523 NMIBC patients treated with transurethral resection and intravesical BCG (2006–2022). After quality filtering, 327 patients (2256 cystoscopy images plus clinical records) were used for classification, and 315 for survival analysis. We developed a multimodal deep learning model (MIBR) combining cystoscopic images and clinical data within a Weibull-based survival framework. Performance was evaluated against image-only, clinical-only, and established risk scores using areas under the receiver operating characteristic curve (AUROC), time-dependent AUROC, and integrated Brier score.

Results: 

For classification, the multimodal model (MIBR-C) achieved an AUROC of 0.84 (95% CI, 0.71–0.96), outperforming the image-only model (0.57), the clinical-only model (0.64), and clinical scores (0.55/0.60/0.65 for European Organization for Research and Treatment of Cancer (EORTC)/Club Urológico Español de Tratamiento Oncológico (CUETO)/European Association of Urology (EAU), respectively;P<0.001). For survival prediction, the multimodal Weibull model (MIBR-W) demonstrated lower error (IBS 0.13 vs. 0.23/0.20/0.24 for EORTC/CUETO/EAU, respectively;P<0.001) and higher time-dependent AUROCs at 10 years compared with Kaplan–Meier-based baselines.

Conclusion: 

In a cohort, uniformly treated with intravesical BCG after surgery, the explainable multimodal deep learning model provided more accurate recurrence prediction and individualized survival trajectories than established clinical scoring systems. These findings highlight the potential of multimodal artificial intelligence models to enhance risk-adapted follow-up and patient counseling in NMIBC.