AI on prostate MRI works best as biopsy triage, not a replacement

Design

Retrospective multicentre, multivendor diagnostic accuracy study

Population

787 men with prostate MRI and histological verification (48.3% with significant cancer)

Primary outcome

Detection of clinically significant prostate cancer versus PI-RADS

Effect

Patient-level sensitivity 97.6% vs 92.6%; AUC 0.80 vs 0.77; lesion-level sensitivity 78.8% vs 88.8%

The question with prostate MRI AI is not whether it can read a scan but where it helps. This multicentre, multivendor study compared a commercial AI risk classifier with radiologist-assigned PI-RADS in 787 men with histological verification, against clinically significant prostate cancer (Grade Group 2 or higher).

At the patient level, AI sensitivity was non-inferior to PI-RADS (97.6% vs 92.6%) and discrimination was slightly higher (area under the curve 0.80 vs 0.77). At the lesion level, though, AI was less sensitive (78.8% vs 88.8%), so it is weaker at precisely localising disease. The value showed in triage: a strategy targeting PI-RADS 3 lesions avoided 18.9% of biopsies while keeping 98.4% sensitivity, and among PI-RADS 3 lesions AI upgraded 93.1% of cancer-positive ones while calling 24.1% of benign ones low risk.

The practical reading is to use AI as decision support for biopsy triage, particularly at equivocal PI-RADS 3 lesions, while the radiologist keeps responsibility for mapping lesions for targeting. It sharpens the biopsy decision rather than replacing the read.

• Multicentre study of 787 men comparing a commercial AI classifier with PI-RADS against Grade Group 2 or higher cancer.
• Patient-level sensitivity was non-inferior (97.6% vs 92.6%) with slightly higher discrimination (AUC 0.80 vs 0.77).
• Lesion-level sensitivity was lower for AI (78.8% vs 88.8%), so localisation is weaker.
• A PI-RADS 3 triage strategy avoided 18.9% of biopsies while keeping 98.4% sensitivity.
• Use AI to support biopsy triage at equivocal lesions, with the radiologist still mapping lesions for targeting.

Why it matters

It reframes the AI-versus-radiologist debate into a concrete role, triaging biopsies at equivocal lesions, rather than a contest the headline numbers imply.

Don't overread it

This was retrospective, and lesion-level sensitivity was inferior to PI-RADS, so AI cannot take over lesion mapping for targeted biopsy.

The statistics, in plain English

Non-inferior patient-level sensitivity means AI did not miss more men with significant cancer overall, but the lower lesion-level sensitivity means it is less reliable at pinpointing where to biopsy, which is why it supports triage rather than 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.

Get Daily Doctor — free

Scan to keep reading on your phone. No account needed to start.