FDA awarded Cognita Imaging a $1.29 million research contract to test whether a panel of large language models can judge the output of another AI model. The 18 month contract started June 22 and runs through the regulatory science group at CDRH. Cognita announced it on September 16.
The work covers 1 million patient exams. Cognita will check how the scoring holds up across patient populations, care settings, imaging hardware and disease types. The question FDA wants answered is which studies still need a radiologist to look.
Cognita was founded in 2024 out of Stanford and sold to Radiology Partners' Mosaic Clinical Technologies in November 2025. Its vision language model for chest X-rays won Breakthrough Device designation in March 2026.
A generative model can produce hundreds of findings per study, which breaks the sensitivity and specificity arithmetic a traditional CADe submission rests on. FDA opened a docket on generative AI devices in August. It is now paying to build the measuring stick those submissions will be held to.