Artificial intelligence in prostate MRI: Comparative diagnostic performance in a high-prevalence cohort.
Authors
Affiliations (8)
Affiliations (8)
- Medical Faculty, University of Oslo, Oslo, Norway.
- Department of Radiology, Akershus Univesity Hospital, Lørenskog, Norway.
- Health Services Research Unit, Akershus University Hospital, Lørenskog, Norway.
- Sunnmære MR-Klinikk, Ålesund, Norway.
- Department of Urology, Akershus University Hospital, Lørenskog, Norway.
- Department of Oncology, Akershus University Hospital, Lørenskog, Norway.
- Department of Applied Mathematics and Theoretical Physics, University of Cambridge, Centre for Mathematical Sciences, Cambridge, UK.
- Institute of Clincal Medicine, University of Oslo, Oslo, Norway.
Abstract
BackgroundArtificial intelligence (AI) is increasingly used in prostate cancer diagnostic workflows but remains insufficiently validated in high-prevalence cohorts often encountered in academic referral centers.PurposeTo compare the diagnostic performance of licensed AI software with routine radiologist readings of prostate MRI, using histopathology as the reference standard.Material and MethodsIn this retrospective study, 1000 patients underwent prostate MRI for suspected prostate cancer (between May 2020 and December 2024), followed by transperineal biopsy; 391 subsequently underwent radical prostatectomy. Diagnostic performance was assessed using sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy across PI-RADS thresholds. Receiver operating characteristic (ROC) analysis, Cohen's kappa for inter-rater agreement, and paired McNemar's test were performed.ResultsIn total, 959 patients (mean age = 69.7 ± 8.4 years) were included; clinically significant prostate cancer (csPCa) was found in 830 (86.5%) patients. AI assigned more cases to PI-RADS 1-2 and fewer to PI-RADS 3 (κ = 0.388; <i>P</i> <.001). At the PI-RADS ≥3 threshold, radiologists showed sensitivity 96.0%, specificity 20.9%, accuracy 85.9%, PPV 88.7%, and NPV 45.0%, compared with 91.3%, 37.2%, 84.0%, 90.3%, and 40.0% for AI (<i>P</i> <.001). ROC performance was comparable (AUC 0.763 vs. 0.739; <i>P</i> = .28). In the prostatectomy subgroup, AI demonstrated a higher false-negative rate (8.4% vs. 4.1%; odds ratio = 3.83, 95% confidence interval [CI] = 1.52-11.51).ConclusionAI showed diagnostic performance comparable to that of radiologists for csPCa detection, with no significant difference in AUC. AI reduced indeterminate PI-RADS 3 scores but missed more significant cancers.