AI read prostate MRI about as well as radiologists across 12 countries — but not at every site

The edition in brief

SCARLET-1 trained an AI system on more than 20,000 prostate MRIs and tested it on 1,974 external patients from 20 cities in 12 countries. Agreement with standard-of-care diagnoses of grade group 2 or higher cancer was 83% to 88%, meeting prespecified interchangeability criteria in all three external datasets, with AUCs of 0.84 to 0.89; but performance varied widely between sites and with image quality. Across 304,455 thyroid ultrasound reports in a US network, TI-RADS language rose from 1.6% to 54.1% between 2017 and 2023, while the share of reports documenting nodules meeting biopsy criteria fell from 23.3% to 8.0%; biopsy rates were unchanged. A locally hosted language model rewrote 360 Dutch primary care X-ray and ultrasound reports into plain language: conclusions met quality criteria in 96%, but 12.5% contained clinically relevant hallucinations. The 2026 ACR Appropriateness Criteria for endometrial cancer rate pelvic MRI usually appropriate for local staging and routine surveillance imaging usually not appropriate for asymptomatic patients treated for low- or intermediate-risk disease.