An interpretable contrast-enhanced CT-based cross-attention transformer model for predicting progression-free survival in muscle-invasive bladder cancer: a multicenter study.

BACKGROUND: Patients with muscle-invasive bladder cancer (MIBC) typically have a dismal prognosis, even after undergoing radical cystectomy (RC). To address this limitation, we constructed and evaluated a preoperative deep learning (DL) model that leverages computed tomography (CT) imaging to stratify MIBC patients according to their risk of postoperative disease progression. METHODS: Data from MIBC patients undergoing RC at one of four institutions were retrospectively analyzed and allocated to training and validation cohorts. A dual-channel information fusion DL model was developed to predict the postoperative risk of progression according to preoperative CT images. Subsequently, a DL-based clinical signature was created by combining clinical features with the DL-derived score. The concordance index (C-index) and area under the curve (AUC) were used to evaluate the DL model's predictive capability, which was compared with the DL-based clinical signature, a clinical model, and a Ki67 model. RESULTS: In the retrospective cohort of 427 patients with MIBC, the DL score demonstrated favorable predictive performance for disease progression in the validation cohort (C-index: 0.774; AUC: 0.785), and generally showed higher predictive performance than the comparator models in the validation cohort. Using an optimal threshold, the patient cohort was categorized into two distinct subgroups, high-risk and low-risk, with differences in the cumulative incidence of progression observed between the two groups, supporting the model's ability to stratify patients according to progression risk. CONCLUSIONS: The utilization of preoperative CT imaging in a DL framework enabled PFS prediction and risk stratification in patients with MIBC, with potential value for individualized postoperative surveillance.