Radiologist versus artificial intelligence in prostate mpMRI: A PI-RADS based diagnostic performance study

Abstract

Aim: To compare radiologist and artificial intelligence (AI) PI-RADS assessments of prostate mpMRI and their performance for clinically significant prostate cancer detection.Materials and Methods: This retrospective single-center study included 460 mpMRI examinations interpreted by radiologists and then analyzed by an AI decision-support system. Patient- and lesion-level PI-RADS scores were compared using weighted Cohen’s kappa and Spearman correlation. Diagnostic performance was assessed in biopsied patients (n=161).Results: Patient-level agreement was moderate-to-good (κ=0.587), with 48.1% exact and 83.5% within-one-category agreement. Lesion-level agreement was moderate (κ=0.442). Adenocarcinoma was found in 50.3% of biopsied patients. For clinically significant cancer (ISUP ≥2), radiologists had higher sensitivity and AI slightly higher specificity at PI-RADS 4–5.Conclusion: AI may serve as a complementary tool in prostate mpMRI interpretation.