Prostate MRI AI matches radiologist agreement in screening datasets across 12 countries

An AI system for reading prostate MRI matched the agreement level of an average radiologist in detecting clinically significant prostate cancer across three external datasets, one diagnostic and two drawn from organised screening programmes, according to the SCARLET-1 study published in Radiology in October 2026¹. The study counted 22,445 examinations in total¹.

What the study found

SCARLET-1 is a retrospective validation study using prostate MRIs from adults with suspected prostate cancer taken between January 2012 and October 2024¹. The AI was trained and tested internally on 20,471 examinations from 19,278 patients in 26 cities across 14 countries¹. External testing then used 1,974 patients from 20 further cities in 12 countries¹. Across the whole study, the 22,445 examinations came from 21,252 men with a median age of 66¹.

The target was Gleason grade group 2 or above, the threshold the authors used for cancer that matters clinically¹. The primary outcome was how often the AI's assessment agreed with the standard-of-care diagnosis, judged against the estimated agreement of an average radiologist¹. The external datasets were PRIME, a diagnostic setting, and STHLM3-MRI and IP1-PROSTAGRAM, both organised screening settings¹.

Agreement was 83.0% in PRIME (395 of 476), 87.9% in STHLM3-MRI (974 of 1,108) and 87.7% in IP1-PROSTAGRAM (342 of 390), meeting the prespecified criteria for diagnostic interchangeability in all three¹. The area under the curve was 0.86 in PRIME, 0.89 in STHLM3-MRI and 0.84 in IP1-PROSTAGRAM, though the confidence interval in the smallest dataset was wide, running from 0.70 to 0.95¹.

Performance held steady across age, risk level, tumour location and ethnicity, with a mean difference in AUC of 0.02¹. It did not hold steady across sites. Of 25 sites, 28% recorded an AUC below 0.80 and 32% an AUC of 0.90 or higher¹. Image quality also mattered, with AUCs ranging from 0.78 to 0.88¹.

What it means for the market

Earlier work had already shown AI systems to be non-inferior to radiologists for detecting prostate cancer on MRI, but the authors note that evidence from diagnostic and population screening settings around the world has been limited¹. SCARLET-1 adds external data from two organised screening datasets as well as a diagnostic one, which is the setting health systems weighing MRI-based screening would care about most¹.

For buyers, the more useful finding may be the spread between sites. A product that clears the interchangeability bar on pooled data can still fall below an AUC of 0.80 at more than a quarter of individual sites, and image quality alone moves the AUC between 0.78 and 0.88¹. That argues for local validation and for attention to scanner and acquisition quality before any trust or screening programme relies on such a tool. The abstract does not name the AI vendor, give costs or report any effect on radiologist workload, so the commercial case remains to be made.

Caveats

The study is retrospective, comparing AI readings with past diagnoses rather than testing the tool in live practice¹. The IP1-PROSTAGRAM dataset was the smallest at 390 patients, and its AUC carried the widest uncertainty¹. Variation across sites and by image quality was substantial¹. The authors conclude that the AI performed comparably to radiology readings across the external diagnostic and screening datasets¹.