Diagnostic performance benchmarks of MRI for prostate cancer: a large-scale cohort study of 33,759 patients

Abstract Objective To evaluate the real-world diagnostic performance of PI-RADS for detecting clinically significant prostate cancer (csPCa) across large-scale worldwide cohorts. Materials and methods This multicenter observational study included 20,429 patients from nine Chinese institutions and a published worldwide meta-analysis dataset comprising 13,330 patients for comparison. All PI-RADS assessments were extracted from original MRI reports produced by clinical radiologists as part of standard diagnostic workflows. Summary receiver operating characteristic curves were constructed using a bivariate random-effects model, and center-level areas under the curve (AUCs) were pooled via random-effects meta-analysis. Subgroup analyses were conducted. Results PI-RADS exhibited overall discriminatory performance for csPCa across nine Chinese cohorts comparable to the META results from 70 studies, though significant inter-center variation was observed. Using a PI-RADS ≥ 3 threshold, pooled sensitivity was 92.6% and specificity was 51.5%. At a PI-RADS ≥ 4 threshold, pooled specificity improved to 78.1%, with a sensitivity of 81.5%. The overall diagnostic performance was consistent across thresholds: The summary ROC (SROC) AUC was 0.903 for PI-RADS ≥ 3 and 0.826 for PI-RADS ≥ 4 in detecting csPCa at the meta-analysis level. Subgroup analyses indicated better performance among older patients (≥ 70 years) and those with higher PSA levels (> 20 ng/mL). PI-RADS v2.1 outperformed v2 or earlier versions in most clinical scenarios. Conclusions This large-scale real-world study provides critical multicenter evidence supporting the use of PI-RADS in clinical practice worldwide, while also revealing substantial variability in scoring and diagnostic performance across institutions. The findings offer an evidence-based overview of PI-RADS implementation in China and worldwide, highlighting both its utility and inconsistencies. Key Points Question What are the real-world PI-RADS diagnostic performance benchmarks for clinically significant prostate cancer detection across diverse Chinese multicenter institutions versus global data? Findings PI-RADS demonstrated good overall performance (SROC AUC 0.903 at a threshold of ≥ 3), with superior discriminatory ability in older patients, higher PSA levels, and v2.1 users. Critical relevance statement This article critically assesses substantial inter-institutional heterogeneity in PI-RADS performance, highlighting that pooled estimates should not replace local calibration and that regionally adapted implementation strategies are essential for advancing clinical radiology practice.