Multicentre evaluation of artificial intelligence risk classification for detection of clinically significant prostate cancer on biparametric MRI.
Authors
Affiliations (9)
Affiliations (9)
- Department of Urology, Vivantes Klinikum Am Urban, Berlin, Germany.
- Prostate-Diagnostic-Centre Berlin, PDZB, Berlin, Germany.
- Department of Urology, Faculty of Health Sciences Brandenburg, Brandenburg Medical School Theodor Fontane, Neuruppin, Germany.
- DeepHealth Inc., Somerville, MA, USA.
- Department of Radiological Sciences, Oncology and Pathology, Sapienza University, Rome, Italy.
- Rolling Oaks Radiology, Thousand Oaks, CA, USA.
- RadNet, Inc., Thousand Oaks, CA, USA.
- Department of Radiology, University College London Hospital NHS Foundation Trust, London, UK. [email protected].
- Division of Surgery & Interventional Science, University College London, London, UK. [email protected].
Abstract
To compare a commercially available artificial intelligence (AI) risk-classification system with PI-RADS for detecting clinically significant prostate cancer (csPCa) on prostate MRI. This retrospective multicentre, multivendor diagnostic accuracy study included men who underwent prostate MRI and histopathologic verification between 2014 and 2025. Index tests were DeepHealth Prostate Suite AI risk category and radiologist-assigned PI-RADS; the reference standard was International Society of Urological Pathology Grade Group ≥ 2. Patient-level non-inferiority testing, receiver operating characteristic analysis, lesion-level free-response receiver operating characteristic analysis, and patient-level bootstrap confidence intervals were used. Workflow support analyses modeled biopsy-avoidance strategies and AI-based risk stratification within PI-RADS 3 lesions. After exclusions, 787 men (median age, 70 years; interquartile range, 64-75) were evaluated; 380 (48.3%) had csPCa. AI sensitivity was non-inferior to PI-RADS at the patient level (97.6% vs 92.6%; p < 0.001) but not at the lesion level (78.8% vs 88.8%; p = 0.98). Patient-level area under the curve was higher for AI than PI-RADS (0.80 vs 0.77; difference, 0.032; 95% confidence interval, 0.001-0.064; p = 0.043). A targeted PI-RADS 3 strategy avoided 18.9% of biopsies while maintaining 98.4% sensitivity. Among PI-RADS 3 lesions, AI upgraded 93.1% of csPCa-positive lesions and assigned a low risk to 24.1% of benign lesions. The AI system was non-inferior for patient-level sensitivity but not lesion-level sensitivity. AI risk classification is best positioned as decision support for biopsy triage, particularly in PI-RADS 3 lesions. Question Can three-tier artificial intelligence risk classification support prostate MRI biopsy triage in men while preserving detection of clinically significant prostate cancer? Findings Artificial intelligence had non-inferior patient-level sensitivity and higher patient-level discrimination than PI-RADS, but lower lesion-level sensitivity at the predefined threshold. Clinical relevance Artificial intelligence risk categories can complement PI-RADS by supporting high-sensitivity biopsy triage, especially for equivocal PI-RADS 3 lesions, while radiologists retain responsibility for lesion mapping.