Synergistic Integration of Radiomics and Deep Features from Multiparametric MRI for Enhanced Prostate Cancer Classification

Accurate classification of clinically significant (CS) versus clinically insignificant (CiS) prostate cancer is critical for treatment decisions.This study investigates the integration of radiomics and deep features from multiparametric MRI (mpMRI) for automated prostate cancer classification.A multimodal deep learning framework with Convolutional Block Attention Module processes three optimally selected MRI sequences (T2-weighted, high b-value DWI, and Ktrans) through parallel encoders.Deep features are extracted from the trained framework, while radiomics features are extracted from each sequence and then concatenated across sequences.Six feature configurations (radiomics-only, deep-only, radiomics-PCA, deep-PCA, combined, and combined-PCA) are evaluated using eleven machine learning classifiers on the official ProstateX challenge dataset (330 training, 208 test lesions).The highest performance is achieved by Voting 2 ensemble classifier with combined PCA features: 0.943 AUC (95% CI: 0.908-0.971),89.9% accuracy, 81.2% sensitivity, and 92.5% specificity.Deep features substantially outperformed radiomics features (mean AUC: 0.899 vs 0.829, 8.44% improvement).Combined features with PCA significantly outperformed deep-only features (2.06% improvement, p = 0.032).Cross-modal correlation analysis (mean |r| = 0.14 ± 0.13) provided theoretical validation that radiomics and deep features capture complementary information.This study demonstrates that systematic integration of radiomics and deep features with PCA-based dimensionality reduction achieves superior prostate cancer classification, offering a validated approach for clinical decision support.