Construction and validation of a multimodal MRI-based deep learning model for early differential diagnosis of prostate cancer in the PSA gray zone: a retrospective cohort study

Background: The diagnostic challenges inherent in prostate-specific antigen (PSA) levels between 4-10 ng/mL represent a critical clinical dilemma, with only 25-30% of patients harboring clinically significant prostate cancer, leading to substantial rates of unnecessary biopsies and associated morbidity. Objective: To develop and validate a multimodal convolutional neural network integrating T2-weighted imaging, diffusion-weighted imaging, apparent diffusion coefficient maps, and clinical parameters for enhanced detection of clinically significant prostate cancer in the PSA gray zone. Methods: This retrospective cohort study analyzed 305 patients with PSA levels 4-10 ng/mL who underwent multiparametric MRI and subsequent biopsy confirmation. A novel multimodal CNN architecture based on modified U-Net with ResNet-50 backbone was developed, incorporating comprehensive fusion strategies. Decision curve analysis was performed to evaluate clinical utility across a range of threshold probabilities. Results: The proposed multimodal CNN achieved superior diagnostic performance with an area under the curve of 0.913 (95% CI: 0.851-0.975), sensitivity of 85.3% (71.4-94.2%), specificity of 90.9% (78.3-97.5%), and overall accuracy of 88.5% (78.2-95.1%), significantly outperforming PSA alone (AUC 0.592, p<0.001) and PI-RADS assessment (AUC 0.694, p<0.001). Nested 5-fold cross-validation demonstrated consistent performance across folds (AUC range: 0.891-0.928), while extended bootstrap validation with 5,000 iterations confirmed robust stability (AUC standard deviation: 0.032). Inter-reader agreement between the model and expert radiologists demonstrated excellent concordance (κ=0.871, 95% CI: 0.831-0.911). Decision curve analysis confirmed a consistently superior net benefit for the multimodal CNN across clinically relevant threshold probabilities. Conclusions: The multimodal deep learning approach represents a paradigm shift in non-invasive prostate cancer detection, potentially reducing unnecessary biopsies by 40-50% while maintaining exceptional sensitivity for clinically significant disease. Decision curve analysis substantiates the clinical utility of this approach across a broad range of decision thresholds.