Optimización de la predicción de recurrencia y estratificación de riesgos en el cáncer de próstata mediante un modelo de aprendizaje profundo 2.5D: un estudio multicéntrico basado en RM

High tumor recurrence after surgery remains a significant challenge in managing prostate cancer (PCa). We aimed to develop and validate a 2.5D deep learning model based on a transformer architecture utilizing T2WI, ADC, DWI, and CE-T1WI images for the preoperative prediction of biochemical recurrence (BCR) in PCa, and to further investigate its capability for risk stratification.

A total of 923 PCa patients (10 153 images) who underwent radical prostatectomy (RP) at five tertiary medical centers were retrospectively enrolled, with follow-up completed by September 2024. Among the five evaluated classifiers, ResNet18 was selected as the best-performing backbone for feature extraction. A Transformer-based deep learning (DL) model was developed using preoperative mpMRI data, and a deep learning fusion (DLF) model was constructed by integrating DL scores with weighted clinical variables, and its performance was compared with the traditional clinical risk score (CAPRA), a clinical model (Clinical), an ensemble learning model (Ensemble), and a multiple instance learning model (MIL). Model performance was evaluated using receiver operating characteristic (ROC) curves. Model comparisons were conducted using the DeLong test, decision curve analysis (DCA) and calibration curves were used to assess the clinical utility and calibration of the models. Furthermore, Grad-CAM was used to visualize model attention and improve interpretability.

The DLF model exhibited excellent performance in both the validation and test sets. It achieved an AUC of 0.938 (95% CI: 0.884-0.992) in the internal validation cohort and an AUC of 0.935 (95% CI: 0.900-0.969) in the external test cohort. The DLF model significantly outperformed all unimodal models ( P < 0.05, DeLong test), improving AUC by 0.058-0.253 over the clinical model and 0.189-0.299 over CAPRA. Furthermore, the DLF model enabled effective risk stratification, with high-risk patients showing significantly poorer prognostic outcomes than low-risk patients (P < 0.05). It also demonstrated significant predictive power for recurrence events at multiple time points. Specifically, as measured by time-dependent AUC values, the model's predictive performance at 1, 2, and 3 years was 0.866 (95% CI: 0.810-0.921), 0.867 (95% CI: 0.816-0.918), and 0.879 (95% CI: 0.832-0.926), respectively.

The stacked DLF model demonstrated the capability to predict early postoperative recurrence in patients with PCa and effectively identify high-risk cases, highlighting its potential as a valuable tool for optimizing treatment strategies and postoperative surveillance.