AI-assisted TURP slide review: a multi-reader study of efficiency, diagnostic performance, and human-AI error patterns.

Transurethral resection of the prostate (TURP) specimens represent a low-prevalence, high-volume diagnostic task in which pathologists must identify rare malignant foci within large amounts of predominantly benign tissue. Although artificial intelligence (AI) has shown strong standalone performance in prostate pathology, its real-world impact on TURP diagnostic workflows remains insufficiently characterized. We developed an AI tool for slide-level detection of prostatic adenocarcinoma using 6,535 H&E whole-slide images from biopsies and TURP specimens, with feature extraction by UNI, classification through a CLAM framework, and presentation on a custom OpenSeadragon-based viewer. Five readers with different experience levels evaluated 102 TURP slides in two sessions, first unaided and then AI-assisted after an 8-week washout period. Primary outcome was per-slide review time; secondary outcome was diagnostic accuracy. AI assistance significantly reduced review time for all readers, with mean reductions ranging from 29.4% to 53.8%. In mixed-effects modeling, AI was associated with a mean reduction of 34.5 s per slide (95% CI, 30.0-39.0; p < 0.001). Time savings were greatest among less experienced readers. Overall diagnostic accuracy increased from 95.3% in the unaided setting to 98.2% with AI assistance. Non-inferiority was confirmed for all readers, and accuracy even improved significantly in the least experienced readers. Across all readings, classification errors decreased from 24 unaided to 9 aided. AI assistance improved TURP slide review efficiency while maintaining, and in some readers improving, diagnostic accuracy. These findings support AI as a workflow-oriented decision-support tool in TURP pathology, particularly for low-prevalence, high-volume diagnostic settings.