Multicentre evaluation of artificial intelligence risk classification for detection of clinically significant prostate cancer on biparametric MRI
European Radiology, 2026
Abstract
Introduction: Radiologists rely on the PI-RADS scoring system to assess prostate MRI and guide biopsy decisions, but equivocal scores — particularly PI-RADS 3 — remain a persistent clinical challenge. This study examined whether a commercially available three-tier AI risk classification system (QUANTIB® Prostate, DeepHealth) could match or exceed PI-RADS performance in identifying clinically significant prostate cancer (csPCa), with a focus on its potential to support biopsy triage decisions.
Methods: This retrospective, multicentre, multivendor diagnostic accuracy study enrolled 787 men (median age 70 years) who underwent prostate MRI and subsequent histopathological verification between 2014 and 2025. Clinically significant prostate cancer was defined as International Society of Urological Pathology Grade Group ≥2. The AI risk category assigned by the DeepHealth Prostate Suite was compared against radiologist-assigned PI-RADS scores using patient-level non-inferiority testing, ROC analysis, and lesion-level free-response ROC analysis. Additional workflow analyses modelled biopsy-avoidance strategies and AI-based risk stratification within the PI-RADS 3 subgroup.
Results: Of the 787 men evaluated, 380 (48.3%) had confirmed csPCa. At the patient level, AI sensitivity was non-inferior to PI-RADS (97.6% vs 92.6%; p < 0.001), and the AI system achieved a higher patient-level area under the curve (0.80 vs 0.77; difference 0.032; 95% CI, 0.001–0.064; p = 0.043). However, at the lesion level, AI sensitivity did not meet the non-inferiority threshold (78.8% vs 88.8%; p = 0.98). Notably, applying a targeted AI-assisted strategy within PI-RADS 3 lesions enabled avoidance of 18.9% of biopsies while preserving 98.4% sensitivity. Among PI-RADS 3 lesions specifically, the AI system upgraded 93.1% of csPCa-positive cases and assigned low risk to 24.1% of benign lesions.
Conclusion: The AI risk classification system demonstrated non-inferior patient-level sensitivity and superior overall discrimination compared with PI-RADS, positioning it as an effective decision-support tool for biopsy triage — particularly for the equivocal PI-RADS 3 category. Lesion-level performance was lower than PI-RADS at the predefined threshold, reinforcing that radiologists should retain responsibility for lesion mapping, with AI serving as a complement rather than a replacement.
