2. UroToday Home
  3. Recent Abstracts
  4. Urologic Oncology
  5. Bladder Cancer

A Deep Learning Model to Predict Treatment Response and Recurrence from Whole Slide Images in Non-Muscle-Invasive Bladder Cancer - Expert Commentary

Non-muscle-invasive bladder cancer (NMIBC) is characterized by high recurrence rates, with up to 40% of patients experiencing recurrence within one year after treatment. A recent study developed and validated a deep learning-based early recurrence predictive model (ERPM) and a treatment response predictive model (TRPM) using whole slide images to improve clinical decision-making for NMIBC patients.

The investigators assembled a cohort of 1,275 patients with pathologically confirmed NMIBC from five hospitals. They collected 4,395 whole slide images with hematoxylin and eosin (H&E) staining and immunohistochemistry (IHC) staining (including P53, CK20, and Ki67). The ERPM was developed using multi-instance and ensemble learning approaches to predict early recurrence within 2 years. The TRPM was designed to predict Bacillus Calmette-Guérin (BCG) therapy response. The performance was evaluated through area under the curve (AUC) analysis, with patients from one hospital assigned to training and internal validation cohorts and patients from four other hospitals serving as external validation cohorts.

The ERPM demonstrated superior performance compared to the clinical and H&E-based models in predicting early recurrence. In the internal validation cohort, the ERPM achieved an AUC of 0.837 versus 0.645 for the clinical model and 0.737 for the H&E-based model. In external validation cohorts, the ERPM maintained robust performance with AUCs ranging from 0.761 to 0.802. The model significantly stratified recurrence-free survival with a hazard ratio of 4.50 (95% CI 3.10-6.53, p < 0.0001). The TRPM also performed well in predicting BCG-unresponsive NMIBC with 84.1% accuracy. IHC staining significantly enhanced the model's predictive capability, with P53 staining providing the most substantial improvement.

Deep learning models are showing significant promise in predicting outcomes for patients with bladder cancer. The ERPM and TRPM show promising potential for guiding personalized management of NMIBC patients. By stratifying patients based on recurrence risk and treatment response, these models could help optimize treatment strategies and follow-up strategies. Future validation of these approaches prospectively in diverse populations is needed. Furthermore, studies showing multimodal integration of H&E, transcriptomic, genomic and clinical data will be critical to improve predictive performance.

Written by: Bishoy M. Faltas, MD, Director of Bladder Cancer Research, Englander Institute for Precision Medicine, Weill Cornell Medicine

References:

  1. Jiang F, Hong G, Zeng H, Lin Z, Liu Y, Xu A, Shen R, Xie Y, Luo Y, Wang Y, Zhu M, Yang H, Wang H, Huang S, Chen R, Lin T, Wu S. Deep learning-based model for prediction of early recurrence and therapy response on whole slide images in non-muscle-invasive bladder cancer: a retrospective, multicentre study. eClinicalMedicine. 2025;81:103125. https://doi.org/10.1016/j.eclinm.2025.103125.
Read the Abstract