Transformer-Based Deep Learning Integrating Multi-Regional MRI and Clinical Features for Predicting Biochemical Recurrence After Radical Prostatectomy: A Multicenter Study.
Authors
Affiliations (4)
Affiliations (4)
- Department of Radiology, The Third Affiliated Hospital, Southern Medical University, Guangzhou, Guangdong, China.
- Department of Ultrasound, Huizhou Central People's Hospital, Huizhou, Guangdong, China.
- Department of Statistics and Actuarial Science, School of Computing and Data Science, The University of Hong Kong, Hong Kong, China.
- Department of Radiology, Zhujiang Hospital of Southern Medical University, Guangzhou, Guangdong, China.
Abstract
Existing deep learning models for predicting biochemical recurrence (BCR) after radical prostatectomy often ignore the biological context beyond the tumor. This study aims to develop and validate a transformer-based model combining multi-regional magnetic resonance imaging (MRI) features and clinical data to improve BCR prediction after radical prostatectomy in prostate cancer (PCa). We enrolled 232 Pca patients from three medical centers between 2013 and 2020. Multi-parametric MRI sequences, including T2-weighted fat-suppressed and diffusion-weighted imaging, were collected before surgery. Patients from centers 1 and 2 were assigned to a discovery cohort (training: n = 146; internal validation: n = 36), and patients from center 3 (n = 50) formed an external test cohort. A transformer-based deep learning (DL) model was developed using imaging features from intratumoral, peritumoral, and periprostatic regions, combined with clinical variables. Model performance was evaluated using the area under the curve (AUC), accuracy, sensitivity, and specificity with 95% confidence intervals (CI), and prognostic value was assessed via Kaplan-Meier (K-M) survival analysis. The combined three-region MRI-based model demonstrated superior predictive performance for BCR risk across all sets, with an AUC of 0.801 (0.663-0.920) on the external test set and an AUC of 0.799 (0.715-0.872) on the validation set. Notably, the integrated DL model, which combined MRI data with clinical characteristics, outperformed the MRI-based and clinical models. The three-region with clinical combined model improved AUCs from 0.801 (95% CI: 0.663-0.920) to 0.863 (95% CI: 0.697-0.975) on the external test set and from 0.799 (95% CI: 0.715-0.872) to 0.832 (95% CI: 0.746-0.904) on the validation set. K-M curve analysis revealed that the integrated model provided significantly better predictions than the MRI-based and clinical models (p < 0.0001). By combining intratumoral, peritumoral, and periprostatic regions from multi-parametric MRI with clinical data, the transformer-based DL model provides a multimodal approach for predicting BCR in PCa patients.