Can AI read a prostate MRI? The SCARLET-1 study

What was shown

The SCARLET-1 study, published in Radiology on 13 October 2026, tested an artificial intelligence (AI) system called PI-CAI-2B, which reads prostate magnetic resonance imaging (MRI) and looks for clinically significant cancer, that is, ISUP grade 2 or higher: the cancer worth diagnosing and treating.

It is a large study: 22,445 MRI scans from 21,252 patients, from 46 cities in 22 countries, including centres in Spain. Most scans were used to train the system and test it within the same centres. The most relevant part is the external validation: 1,974 patients from centres the AI had never "seen".

This external validation covered two different settings:

• Diagnosis (men with suspected cancer, PRIME study): discrimination (AUC) of 0.86.
• Organised population screening (STHLM3-MRI and IP1-PROSTAGRAM): AUC 0.89 and 0.84.

An AUC of 1 would be perfect, and 0.5 is no better than tossing a coin. The primary outcome was something else: the agreement between the AI and the standard-of-care diagnosis. It was comparable to the estimated agreement of an average radiologist in all three external groups. Performance was similar across age, risk, tumour location and ethnicity.

What it does not show

A good result in a study is not the same as a benefit for patients. SCARLET-1 has limitations worth knowing:

• It is retrospective: it analyses scans that had already been done. It does not measure what happens when AI actually takes part in decisions.
• The reference was not always biopsy. In some cases it was radiologist consensus. A tumour that nobody saw may therefore go uncounted (verification bias).
• The comparator is an estimated average radiologist, not radiologists reading the same images as the AI.
• Performance varies by site: in 28% of 25 sites the AUC was below 0.80. It also varies with image quality (AUC 0.78 to 0.88).

Nor does it show that AI can replace the radiologist, or that an MRI read by software is enough to decide on a biopsy.

What changes in practice

For now, nothing. My reading: promising, but it does not change practice yet.

There is an important nuance: the AI threshold is chosen according to its use. In screening, the aim is to avoid false alarms in healthy men. In diagnosis, the aim is not to miss tumours. Changing the threshold shifts the balance between biopsies avoided and cancers detected, so there is no single "AI result".

Before it is used in clinical practice, prospective studies with clinical outcomes are needed. And each centre would have to validate it on its own MRI scans, because the study itself shows that performance depends on the site and on image quality.

If your PSA is raised, the usual pathway does not change: I explain it in High PSA: what it means and what to do and in Prostate fusion biopsy. For another AI software for MRI, see AI in prostate MRI: what QP-Prostate adds.

Source

Saha A, Bosma JS, et al. Artificial Intelligence for Prostate Cancer Detection and Screening on MRI in Global Populations: The SCARLET-1 Study. Radiology 2026;321(1):e262081. Published online 13 October 2026. PMID 42867371 ↗. Open access: doi:10.1148/radiol.262081 ↗.

Infographic