Artificial Intelligence in Prostate MRI: Comparison of an AI-Based Software and an Experienced Radiologist for Detecting Clinically Significant Prostate Cancer

TL;DR

This study compares an AI-based software with an experienced radiologist for detecting significant prostate cancer on MRI and finds they perform similarly, with AI showing potential as a supportive tool.

Contribution

Demonstrates that AI-based software can match an expert radiologist in detecting clinically significant prostate cancer on MRI.

Findings

AI software and the radiologist showed comparable sensitivity and specificity in detecting clinically significant prostate cancer.

The AI detected more lesions, especially in the transition zone, but should be used as a support tool rather than a replacement for radiologists.

Both methods had extremely low negative likelihood ratios, indicating strong ability to rule out significant cancer.

Abstract

Prostate cancer is one of the most common cancers in men, and magnetic resonance imaging (MRI) plays a key role in identifying tumors that require treatment. However, interpreting prostate MRI requires experience, and results may vary between readers. Artificial intelligence tools have been developed to assist radiologists, but their real clinical value is still being evaluated. In this study, we compared an artificial intelligence–based software with an experienced radiologist in detecting clinically significant prostate cancer on MRI, using biopsy results as reference. We found that the software performed similarly to the expert reader, particularly in ruling out significant cancer when MRI findings were negative. Although the software detected more suspicious areas, it should be considered a support tool rather than a replacement for radiologists. These findings suggest that artificial intelligence could help improve consistency in prostate MRI interpretation and support less-experienced readers, with potential impact on future research and clinical practice.

Background: Multiparametric MRI is central to detecting clinically significant prostate cancer (csPCa), but diagnostic accuracy depends on reader experience. Artificial intelligence (AI) tools may support prostate MRI interpretation and reduce inter-reader variability. This study compared the detection rate of a trial, non-commercial version an AI-based software (PAROS) with that of an experienced radiologist. Methods: This retrospective single-center study included 150 patients who underwent prostate MRI followed by combined systematic and MRI-targeted transperineal biopsy. MRI examinations were interpreted by an experienced radiologist according to PI-RADS v2.1 and independently analyzed using a precommercial trial version of PAROS operating on biparametric MRI. Histopathology served as the reference standard. Detection rate was evaluated using sensitivity, specificity, and positive and negative likelihood ratios (PLR and NLR) at PI-RADS thresholds ≥3 and ≥4. Results: CsPCa was present in 63.3% of patients. At both PI-RADS thresholds, PAROS and the radiologist showed comparable sensitivity and specificity, wuth extremely low NLRs, indicating excellent rule-out capability. PLRs were modest and similar at PI-RADS ≥ 3 (1.26 vs. 1.42) and 1.88 for both at PI-RADS ≥ 4. PAROS detected more lesions, particularly in the transition zone. Conclusions: PAROS achieved csPCa detection comparable to an experienced radiologist, supporting its role as a decision-support tool in prostate MRI interpretation.

Figures