Evaluation of the Effectiveness and Safety of CascadeDiagnose Prostate MRI: A Multi-Center Clinical Study

This study addresses a clinically important gap in prostate cancer diagnosis: although multiparametric MRI is central to detection, PI-RADS interpretation is subject to substantial inter-reader variability and often yields suboptimal discrimination between clinically significant cancer and benign mimics. CascadeDiagnose Prostate MRI is a cascaded deep-learning system that first detects and segments prostatic lesions on T2-weighted, diffusion-weighted, and apparent diffusion coefficient sequences, and then performs malignant-benign classification and Gleason-based risk stratification. The system is designed to provide interpretable, traceable outputs-including sequence and slice references, zonal localization, confidence scores, Grad-CAM heatmaps, and segmentation masks-and to return indeterminate or rejection results when image quality, data integrity, or model confidence is insufficient, rather than forcing a diagnosis. The study will evaluate this system across six tertiary hospitals with diverse MRI vendors, field strengths, and radiologist experience, using both retrospective and prospective validation. A central methodological feature is the Guangxi-wide medical-sharing distributed trusted computing network. Each hospital operates a local edge node; raw MRI, clinical, and pathological data remain within the hospital and are never centrally uploaded. Only de-identified, SM4-encrypted intermediate features and gradient parameters are transmitted to a regional aggregation node for federated model collaboration and multi-center statistical summarization, implementing the principle of "don't move data; move computing power." In the prospective phase, AI analysis and human reading are performed under a locked system version, with pre-specified rules for clinician access to AI outputs. Reading sessions include AI-alone, radiologist-alone, and AI-assisted interpretations to assess diagnostic performance, workflow efficiency, and safety in a real-world setting. The study incorporates rigorous quality control, including local image-quality review, anti-tampering receipts, cross-node audit trails, and full traceability of data, model, threshold, inference, and manual-review versions. It is intended to generate regulatory-grade evidence on diagnostic accuracy, generalization, rejection behavior, and clinical utility while demonstrating a compliant, privacy-preserving infrastructure for multi-center evaluation of healthcare AI.