An interpretable CT-based deep learning model for predicting overall survival in patients with bladder cancer: a...

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Prognosticating bladder cancer survival is difficult with current treatments.
• Accurate prediction of overall survival is crucial for guiding clinical decisions.

Purpose of the Study:

• To develop and validate an interpretable deep learning model (BCDL) for predicting bladder cancer patient survival using preoperative CT scans.
• To assess the BCDL model's performance against existing models and identify key predictive features.

Main Methods:

• A deep learning model was trained on 765 patients and validated on three independent cohorts (438, 181, 72).
• The SHapley Additive exPlanation (SHAP) method was used to interpret model predictions by identifying pixel-level features.
• Patients were stratified into high- and low-risk groups based on BCDL scores.

Main Results:

• The BCDL model demonstrated superior performance in predicting survival risk.
• Adjuvant therapy significantly improved overall survival in high-risk patients (p=0.028) and low-risk women (p=0.046).
• RNA sequencing revealed distinct gene expression patterns, an immunosuppressive microenvironment in high-risk patients, and altered microbial composition.

Conclusions:

• The interpretable BCDL model accurately predicts bladder cancer survival risk from CT scans.
• The model supports personalized treatment strategies by stratifying patients and identifying those who benefit from adjuvant therapy.
• Findings highlight the potential of AI in enhancing clinical decision-making for bladder cancer management.