European Radiology | 2026
Karsten Guenzel; Maarten G. Poirot; Almar van Loon; Emanuele Messina; Martina Pecoraro; Valeria Panebianco; Robert Princenthal; Francesco Giganti
Purpose
This study evaluated whether DeepHealth Prostate Suite’s artificial intelligence (AI) risk classification achieved non-inferior sensitivity to radiologist-assigned PI-RADS for detecting clinically significant prostate cancer on biparametric MRI, at both the patient and lesion levels. It also explored whether combining AI risk categories with PI-RADS could support biopsy decisions, particularly for equivocal PI-RADS 3 findings.
Methods and Materials
This retrospective, multicenter, multivendor study included 787 men from Germany, the Netherlands, and the United States who underwent prostate MRI and histopathologic verification between December 2014 and January 2025. Of these, 380 (48.3%) had csPCa, defined as International Society of Urological Pathology Grade Group ≥2.
The AI analyzed biparametric MRI sequences and assigned Low, Medium, or High risk categories. Investigators compared AI performance with PI-RADS at the patient and lesion levels, using biopsy or prostatectomy histopathology as the reference standard. Additional analyses modeled biopsy strategies combining AI risk categories with PI-RADS.
Results
At the patient level, AI achieved 97.6% sensitivity compared with 92.6% for PI-RADS, meeting the prespecified criterion for non-inferiority (P < .001). These results used thresholds of AI ≥Medium and PI-RADS ≥4; specificity was 19.7% and 48.4%, respectively. The continuous AI risk score also showed higher patient-level discrimination than PI-RADS, with an area under the curve of 0.80 versus 0.77 (P = .043).
A modeled strategy recommending biopsy for PI-RADS ≥4, or PI-RADS 3 with AI ≥Medium, would have avoided biopsy in 18.9% of the study cohort while maintaining 98.4% sensitivity. Among 29 PI-RADS 3 lesions containing csPCa, AI classified 27 (93.1%) as Medium or High risk.
At the lesion level, AI sensitivity was lower than PI-RADS (78.8% versus 88.8%), and non-inferiority was not demonstrated.
Conclusion
DeepHealth Prostate Suite achieved non-inferior patient-level sensitivity and higher overall patient-level discrimination than PI-RADS in this international study. The findings support AI risk classification as a complementary tool for biopsy decision-making, particularly in equivocal PI-RADS 3 cases.
Biopsy reductions were modeled retrospectively and require prospective validation. Lower lesion-level sensitivity underscores the continued importance of radiologist interpretation and comprehensive lesion assessment.
Read full article