The U.S. Food and Drug Administration awarded Cognita Imaging a $1.29 million research contract to test a new approach for evaluating medical reports produced by artificial intelligence systems in radiology. The contract began on June 22, 2026, and runs for 18 months. It examines whether a “panel” of large language models can be used to review the output of another model while identifying cases that require intervention by a radiologist.
The project is being conducted with the regulatory science team of the FDA’s Center for Devices and Radiological Health. Cognita will build the evaluation framework and test it on one million patient examinations within a large U.S. dataset, comparing performance across different patient groups, care settings, imaging devices, and diseases, including rare conditions.
The Open-Ended Evaluation Problem
Most FDA-authorized artificial intelligence tools in radiology are evaluated for a specific task, such as detecting a blood clot in a head CT scan or identifying a collapsed lung on a chest X-ray. Newer models, by contrast, seek to produce a complete radiology report, making the evaluation scope much broader; a report can include hundreds of potential findings rather than one or two.
Akshay Chaudhry, Cognita’s co-founder and an associate professor of radiology and biomedical data science at Stanford University, said that evaluating these outputs through the ongoing use of panels of radiologists becomes more difficult and time-consuming.
How Will the Results Be Reviewed?
Cognita will use several language models to review reports from radiology models. When clinically significant differences emerge, radiologists will intervene to determine the source of the problem: the AI-generated report, the “panel” of language models, or the original report written by the radiologist.
The project focuses on two known risks of generative artificial intelligence: hallucinations, meaning the inclusion of information that appears plausible but is incorrect, and omissions, meaning the failure to mention a finding that should have been included. The methodology aims to measure the rate of each and determine whether its effect is clinically significant.
Why Does This Matter?
The project’s regulatory value does not lie in approving a new tool, because the system under study has not received FDA authorization. More importantly, it tests whether methods for evaluating generative artificial intelligence can be expanded without sacrificing human review in sensitive cases. If the methodology proves valid, it could provide the regulator with a more practical framework for evaluating tools that produce open-ended text.
Cognita will deliver to the FDA a report on the findings and limitations, the software code, guidance for building panels of language models, comparisons between large and small validation sets, and radiologist-reviewed discrepancy studies. However, the contract does not predetermine that the judgments of language models will replace experts; disagreements with clinical impact will still require human review, and the validity of the methodology across different devices, patients, and diseases is among the questions the test is intended to answer.
A Separate Path for a Vision-and-Language Model
Cognita is also working with the FDA on a vision-and-language model that interprets chest X-ray images and generates reports reviewed by a radiologist. Earlier in 2026, this device received FDA “Breakthrough” designation, but it has not yet been authorized. Chaudhry emphasized that this path is separate from the research contract and that the company and the regulator are still determining what type of clinical study is needed to measure accuracy, benefits, and risks.
Cognita was founded in 2024 and acquired by Radiology Partners in 2025. This work comes amid a shortage of radiologists in the United States, which is associated with longer working hours and a backlog of cases. The project does not yet provide an approved solution to this shortage, but it addresses one of the practical obstacles to evaluating tools that might be used to assist with report preparation.