- Design
- Retrospective multicentre, multivendor diagnostic accuracy study
- Population
- 787 men with prostate MRI and histological verification
- Primary outcome
- Detection of Grade Group ≥2 prostate cancer
- Effect
- Patient-level sensitivity 97.6% vs 92.6%; 18.9% of PI-RADS 3 biopsies avoidable at 98.4% sensitivity
This retrospective multicentre, multivendor study compared a commercially available AI risk classification (DeepHealth Prostate Suite) with radiologist PI-RADS in 787 men with prostate MRI and histological verification; 48.3% had clinically significant cancer (Grade Group 2 or higher).
At patient level, AI sensitivity was non-inferior to PI-RADS (97.6% vs 92.6%), and discrimination was slightly higher (AUC 0.80 vs 0.77). At lesion level, AI was not non-inferior (78.8% vs 88.8%). Applied to PI-RADS 3 cases, an AI-guided strategy would have avoided 18.9% of biopsies while keeping 98.4% sensitivity. Among PI-RADS 3 lesions, AI upgraded 93.1% of those that proved to be significant cancer and called 24.1% of benign ones low risk.
The finding lands exactly on the category that causes most unnecessary biopsies. It supports AI as a second opinion for PI-RADS 3 triage, with the radiologist still mapping lesions for targeting, given AI's weaker lesion-level performance. The cancer prevalence here (48%) is higher than in most referral populations, and the biopsy-avoidance figure is modelled, not observed.
- Consider AI risk classification as decision support for PI-RADS 3 lesions where it is available and validated locally.
- Keep radiologist lesion mapping for targeted biopsy; AI missed more individual lesions than PI-RADS.
- Combine AI risk with PSA density and clinical factors rather than using it alone to cancel a biopsy.
- Audit outcomes locally; the modelled biopsy savings came from a high-prevalence cohort.
Why it matters
PI-RADS 3 drives much of the unnecessary prostate biopsy workload.
Don't overread it
Retrospective, high-prevalence cohort; the 18.9% biopsy reduction is modelled, not prospectively observed.
The statistics, in plain English
Non-inferior means AI's patient-level sensitivity was not worse than PI-RADS by more than a preset margin. The AUC difference of 0.032 (0.001 to 0.064) is statistically significant but small. Lesion-level sensitivity 10 points lower matters for biopsy targeting.
Read the rest in the app
You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.
Scan to keep reading on your phone. No account needed to start.
Tomorrow morning, before your first patient
One edition a day for radiology, written by the desk, every claim tied to its paper. Six minutes.
Get the app — free