Emergency departments lean on right upper quadrant ultrasounds because they are fast, painless and cheap. Reading them is not. A Colorado team thinks a language model can close that gap.
Researchers at the University of Colorado Anschutz School of Medicine built a vision-language agent that reads these scans, names the likely condition and drafts a preliminary report. The work appears in Nature Communications.
Unlike an image-only model, the system pairs the ultrasound frames with the patient’s lab results. That context matters, since elevated liver enzymes or bilirubin often trigger the order in the first place.
Harrison Bai, the senior author, framed the tool as a colleague rather than a competitor. In his words it is not an attempt to build an AI radiologist, but a way to hand radiologists a head start on reporting.
Speed carries real weight in this corner of the abdomen. Acute cholecystitis, an inflamed gallbladder, can demand surgery without delay. An agent that flags those cases early can compress the route from scan to operating table.
The accuracy headline was 99.6% in the reported study. The researchers also stress consistency. Right upper quadrant images carry substantial noise and artifact, and skill varies widely among the people reading them.
The backdrop is a radiology workforce under strain. Imaging demand keeps climbing as the population ages, and radiologists were exiting the field at twice the 2014 rate by 2022, according to the paper. Medicare reimbursement has tightened at the same time.
Bai argues that time saved on reporting is the adoption lever. Clinicians who feel overrun respond to a tool that shortens paperwork rather than one that promises better diagnoses.
The agent generates and the radiologist reviews. That division keeps a licensed human in the loop for every cholecystitis call, every suspected stone and every incidental finding.
