Clinical-imaging fusion model for risk assessment of clinically significant prostate cancer

Abstract

Background Prostate cancer is the second most commonly diagnosed malignancy among men worldwide. Its biological heterogeneity challenges conventional imaging-based risk assessment and complicates biopsy decision-making. Although advanced artificial intelligence may improve risk stratification, medical imaging applications are often limited by small sample sizes. Vision foundation models, with strong transferability in data-constrained settings, offer a promising approach to improving prostate cancer risk assessment, guiding personalized treatment, and reducing overdiagnosis. Methods A robust transfer learning framework based on prostate magnetic resonance imaging (MRI), termed robust transfer learning model (RTLM), was developed to enable the non-invasive risk assessment of clinically significant prostate cancer (csPCa). RTLM employs a feature-matching transfer strategy to adaptively capture task-relevant abstract knowledge from vision foundation models, thereby enhancing the feature representation capability and robustness of convolutional neural networks. Model performance was evaluated on csPCa MRI data, and a multi-task evaluation was conducted on PD-L1 expression prediction in patients with non-small cell lung cancer (NSCLC) to assess the generalizability of the proposed framework. Subsequently, features extracted by the RTLM were used to construct a deep learning signature (DLS), which was then integrated with key clinical variables, including age, serum total PSA level, PSA density, prostate volume, and PI-RADS score, to build the clinical-imaging fusion model (CIFM) for csPCa risk assessment. Diagnostic performance was evaluated using the area under the curve (AUC), decision curve analysis (DCA), integrated discrimination improvement (IDI), and net reclassification improvement (NRI). Results The CIFM achieved AUCs of 0.918, 0.890, 0.828, and 0.852 in the training cohort (n = 585), test cohort (n = 310), and two external validation cohorts (n = 510 and n = 94), respectively. Compared with the clinical model and the RTLM, CIFM showed significant improvements in both IDI and NRI (all p < 0.05). Decision curve analysis further demonstrated that CIFM provided greater net clinical benefit. Conclusion By effectively leveraging knowledge representations from vision foundation models, RTLM enhances the feature learning capability and robustness of CNNs in small-sample medical imaging tasks. The CIFM, which integrates clinical indicators, demonstrates good accuracy in csPCa risk assessment, suggesting potential generalizability.