RESEARCH SUMMARY

Multinational Study: AI Boosts Prostate Cancer Detection Accuracy in MRI Analysis

Anthony Calabro, MA

Key Highlights

  • Artificial intelligence (AI) assistance increased diagnostic accuracy for detecting clinically significant prostate cancer
  • Sensitivity improved from 94.3% to 96.8% and specificity from 46.7% to 50.1% with AI support.
  • Nonexpert readers gained more from AI assistance than experts, narrowing performance gaps.
  • Stand-alone AI outperformed both unassisted and assisted human readers, suggesting future integration potential.

In a large-scale, international observer study involving 61 radiologists from 53 centers in 17 countries, concurrent use of a scientifically validated artificial intelligence (AI) system significantly improved the diagnosis of clinically significant prostate cancer (csPCa) in biparametric MRI examinations.

Readers assisted by AI showed a statistically significant increase in diagnostic accuracy, sensitivity, and specificity when compared with their unassisted assessments. Notably, nonexpert readers experienced greater performance gains with AI, helping to close the gap with expert-level diagnostic proficiency.

MRI is increasingly adopted worldwide for prostate evaluation but remains susceptible to interreader variability and dependence on high-level expertise. The Prostate Imaging-Cancer AI (PI-CAI) study previously demonstrated strong standalone performance of AI in csPCa detection. This follow-up study aimed to determine whether AI could assist human readers in real-time and improve diagnostic outcomes in a controlled crossover design.

The study included 360 prostate MRI examinations from four European centers and evaluated reader performance both with and without AI support. Readers provided PI-RADS scores and patient-level suspicion ratings on a 0-100 scale. The primary endpoint was the area under the receiver operating characteristic curve (AUROC), with additional analysis of sensitivity and specificity at a PI-RADS threshold ≥ 3.

AI assistance increased AUROC by 3.3% (from 0.882 to 0.916; 95% CI, 1.8%-4.9%; P < .001). Sensitivity improved by 2.5% (from 94.3% to 96.8%; P < .001), while specificity rose by 3.4% (from 46.7% to 50.1%; P = .01). AI also facilitated a reduction in false positives and added three true-positive diagnoses. In 33% of assessments, readers changed their patient-level scores when aided by AI, with 8% upgraded and 9% downgraded from initial impressions.

Stratified analyses revealed that nonexpert readers benefited more from AI, with AUROC gains of 0.053 versus 0.018 for experts. Sensitivity and specificity improved more substantially among nonexperts, indicating that AI may be particularly valuable in settings lacking high-volume prostate MRI expertise.

This study has limitations, including its retrospective design, cohort derivation from a prior AI study, use of a controlled viewing platform rather than native workstations, and lack of workflow efficiency data. Additionally, not all patients with negative MRI received biopsy confirmation, limiting histopathologic validation in some cases.

“The findings of this diagnostic study suggest the potential of AI assistance in improving csPCa diagnosis when compared with unassisted assessments of biparametric MRI, with statistically significant improvements observed across AUROC, sensitivity, and specificity at a PI-RADS score of 3 or more,” the study authors concluded.

Reference
Twilt JJ, Saha A, Bosma JS, et al. AI-Assisted vs Unassisted Identification of Prostate Cancer in Magnetic Resonance Images. JAMA Netw Open. 2025;8(6):e2515672. Published 2025 Jun 2. doi:10.1001/jamanetworkopen.2025.15672