Abstract
Abstract
Accurate classification of clinically significant (CS) versus clinically insignificant (CiS) prostate cancer is critical for treatment planning, yet clinical adoption of AI-based diagnostic systems remains limited by two fundamental barriers: achieving clinically meaningful performance with balanced sensitivity/specificity for decision support, and lack of transparent decision-making that clinicians can verify and trust. This study addresses both barriers through clinical validation of a multimodal deep learning framework that achieves clinically useful classification performance while providing interpretable, verifiable predictions. The framework employs three parallel encoders with Convolutional Block Attention Module (CBAM) integration to process complementary mpMRI sequences (T2W, high b-value DWI, and Ktrans) through attention-enhanced feature refinement and multimodal fusion. Evaluated on the ProstateX challenge dataset, the framework achieves 0.91 AUC (95% CI 0.85–0.96) with clinically balanced performance: 79% sensitivity for identifying significant cases and 90% specificity for correctly identifying clinically insignificant cases—critical for reducing unnecessary biopsies and interventions. Most importantly, we validate model trustworthiness through systematic integration of four complementary explainable AI techniques—Grad-CAM, SmoothGrad-CAM, CBAM spatial attention visualization, and permutation-based feature importance analysis—demonstrating that model decisions consistently align with established radiological diagnostic criteria. This work provides evidence that attention-enhanced multimodal learning can achieve both clinically meaningful classification performance and transparent decision support, establishing a foundation for trustworthy AI-assisted prostate cancer diagnosis in clinical practice.