Development and validation of a radiomics deep learning signature from MRI-guided transrectal ultrasound combined with ultrasound imaging parameters for prostate cancer prediction

Objective This study aimed to develop and validate an integrated radiomics deep learning signature combining MRI-guided transrectal ultrasound (TRUS) and contrast-enhanced ultrasound (CEUS) parameters for improved prediction of prostate cancer (PCa). Methods This bicentric retrospective study enrolled 443 patients with suspected PCa confirmed by histopathology. Each patient underwent multiparametric MRI and TRUS-guided CEUS. Radiomic and deep learning features were extracted from B-mode ultrasound images using an AI-based platform. Feature selection was performed using statistical and regression methods. Machine learning classifiers were developed and merged with clinical parameters into a combined model. Performance was evaluated via ROC analysis, calibration curves, and decision curve analysis. Results The SVM-based radiomics model achieved an area under the curve (AUC) of 0.936 (95% CI: 0.905–0.966) in the training cohort and 0.823 (95% CI: 0.721–0.925) in the validation cohort. The clinical model alone yielded an AUC of 0.856 (95% CI: 0.812–0.901) in the training cohort and 0.809 (95% CI: 0.708–0.910) in the validation cohort. The integrated radiomics-clinical model demonstrated superior performance, with an AUC of 0.956 (95% CI: 0.931–0.981) in the training cohort and 0.889 (95% CI: 0.813–0.966) in the validation cohort. DCA confirmed the clinical utility of the combined model across a wide range of threshold probabilities. Conclusion The integration of radiomics deep learning features with CEUS parameters and clinical risk factors significantly enhances the accuracy of PCa prediction. This non-invasive approach shows promise for supporting clinical decision-making and reducing unnecessary biopsies.