Clinical validation of an AI-based computer-aided detection system for risk classification of radiologist-identified prostate MRI lesions

AI-based tools for MRI demonstrate high sensitivity for clinically significant prostate cancer (csPCa) detection, yet performance varies across cohorts. We evaluated a commercial AI-based computer-aided detection (AI-CAD) system for csPCa classification in radiologist-identified suspicious lesions in a real-world setting. This retrospective, single-centre study included 113 men with a PI-RADS ≥ 3 index lesion undergoing 3T in-bore MRI-targeted robotic biopsy. Cases were analysed using QP-Prostate ® . Lesion-level diagnostic performance was assessed for csPCa (Gleason ≥ 7) and overall PCa (Gleason ≥ 6) under two positivity thresholds (Moderate + High and High-only). Analyses were stratified by prostate zone; exploratory analyses included PI-RADS 3 lesions and a theoretical simulation of biopsy-sparing. Mean age was 70.1 (SD 7.9) years; csPCa prevalence was 58.4%, and 15.9% of lesions were PI-RADS 3. For csPCa classification, sensitivity and specificity were 0.92 and 0.32 (Moderate + High), and 0.79 and 0.43 (High-only), respectively. Performance was higher in the peripheral zone (PZ) than in the transitional/central zones (TZ + CZ) across both thresholds. For overall PCa, sensitivity and specificity were 0.91 and 0.43 (Moderate + High), and 0.80 and 0.61 (High-only), respectively. In PI-RADS 3 lesions, csPCa sensitivity was 0.33 across thresholds, with specificity improving from 0.60 (Moderate + High) to 0.67 (High-only). In the theoretical biopsy-sparing simulation, 20/113 biopsies (17.7%) would have been withheld with Moderate + High and 34/113 (30.1%) with High-only, with 5/66 (7.6%) and 14/66 (21.2%) csPCa cases missed, respectively. Among patients without csPCa, 15/47 (31.9%) and 20/47 (42.6%) biopsies would have been avoided, respectively. In a challenging cohort of radiologist-identified PI-RADS ≥ 3 lesions, the AI-CAD system maintained high sensitivity for csPCa classification particularly in the PZ, with lower performance in the PI-RADS 3 exploratory subgroup. The theoretical biopsy-sparing simulation showed a threshold-dependent trade-off between potential biopsy withholding and missed csPCa. Prospective validation is required before AI-CAD-informed biopsy avoidance can be considered clinically safe and effective.