Cognita CEO explains next steps for vision language models in radiology

Katie Palmer covers telehealth, clinical artificial intelligence, and the health data economy, with a focus on the impacts of digital healthcare for patients, providers, and businesses. You can reach Katie on Signal at palmer.01.

The largest radiology practice in the United States is relying even more on artificial intelligence. The technology arm of Nashville-based Radiology Partners, which includes more than 4,000 radiologists reading more than 55 million images each year, last month acquired a new AI company for $80 million: Cognita Imaging, a startup founded by Stanford researchers that hopes to win the race to capitalize on basic models in radiology.

By training visual language models on a large number of X-ray images and their written X-ray reports, the hope is that the AI ​​will be able to read an X-ray or CT scan like a radiologist would: not just looking for a predetermined single abnormality, but any finding that seems important. Many existing and new radiology companies have embarked on this goal, despite concerns about whether such a broadly targeted technology can be validated and used safely.

Radiology Partners, which has long served as a large-scale testing ground for AI in radiology, could accelerate this work in unique ways. Over the last year, the company’s technology and AI division, Mosaic Clinical Technologies, incorporated Cognita’s model into a tool called Mosaic Drafting, which analyzes X-rays and CT scans of the head and spits out the text of a preliminary radiology report for a human radiologist to review, edit and approve.