Artificial Intelligence for Prostate Cancer Detection and Screening on MRI in Global Populations: The SCARLET-1 Study

Background Artificial intelligence (AI) systems have demonstrated non-inferiority to radiologists for prostate cancer detection on MRI, but evidence in global diagnostic and population-based screening settings remains limited. Purpose To determine whether an AI system can achieve diagnostic performance comparable to radiology readings for detecting Gleason grade group ≥2 prostate cancer on MRI in external diagnostic and screening datasets. Materials and Methods In the SCARLET-1 retrospective validation study, prostate MRIs from adults with suspected prostate cancer (January 2012-October 2024) were used to evaluate an AI system, with external testing in diagnostic (PRIME) and organized screening settings (STHLM3-MRI and IP1-PROSTAGRAM). The primary outcome was agreement between AI assessments and standard-of-care diagnoses of Gleason grade group ≥2 cancer, with diagnostic interchangeability defined relative to the estimated agreement of an average radiologist. Secondary outcomes included area under the receiver operating characteristic curve (AUC) across datasets and subgroups. Results 22,445 examinations (21,252 male patients; median age, 66 years [IQR 60-71]) were included: 20,471 prostate MRI examinations (19,278 patients; 26 cities, 14 countries) for AI training and internal testing, and 1974 patients (20 external cities, 12 countries) for external testing. External AUCs were 0.86 (95% CI: 0.83, 0.90) in PRIME, 0.89 (95% CI: 0.86, 0.92) in STHLM3-MRI, and 0.84 (95% CI: 0.70, 0.95) in IP1-PROSTAGRAM. The primary outcome was 83.0% (395 of 476; adjusted 95% CI: 78.8%, 87.0%), 87.9% (974 of 1108; adjusted 95% CI: 85.6%, 90.3%), and 87.7% (342 of 390; adjusted 95% CI: 83.6%, 91.5%), respectively, meeting prespecified interchangeability criteria in all datasets. Subgroup performance was consistent across age, risk, cancer location and ethnicity (mean ΔAUC, 0.02), but varied substantially across 25 sites (28% with AUC<0.80; 32% with AUC ≥0.90) and by image quality (AUC, 0.78-0.88). Conclusion AI demonstrated diagnostic performance comparable to that of radiology readings across external diagnostic and screening datasets. © The Author(s) 2026. Published by the Radiological Society of North America under a CC BY 4.0 license.