AI drafts abdominal ultrasound reports in study, with radiologists still editing

An AI system that reads abdominal ultrasound images alongside laboratory data has produced promising preliminary reports for radiologists to check. The work, described in Nature Communications, targets both diagnosis and the paperwork that follows it. But researchers still need to establish whether the approach improves accuracy, efficiency or patient well-being in real-world emergency care.

Reading the image and the clinical clues

The research focuses on right upper quadrant ultrasounds: scans of the upper-right abdomen that show organs including the liver, gallbladder and right kidney. Doctors often order them for pain in that area or abnormal blood-test results, such as elevated liver enzymes or bilirubin.

Researchers built a vision-language AI agent, meaning a system that works with both images and text. It combines ultrasound images with written clinical laboratory data. Its assigned tasks extend beyond producing a description: it was designed to screen for 16 common diagnoses, estimate a patient’s prognosis and draft a report for review.

There is a practical obstacle behind that design. Ultrasound images can contain visual noise and artifacts, and readers vary in skill and experience. The researchers see AI as a possible way to make interpretations more consistent. Senior author Harrison Bai, a visiting associate professor of radiology at Colorado, also identifies report writing as a time-consuming part of an increasing workload.

A large scan dataset, a smaller report check

The team evaluated the agent on roughly 11,000 ultrasounds from Stanford University, the University of Colorado and Johns Hopkins University. Researchers then manually reviewed 100 AI-generated reports for accuracy. These are different scales of evaluation: the full scan dataset should not be confused with the smaller set of reports checked by hand.

Within that reviewed sample, researchers reported correct diagnosis identification 99.6% of the time. That figure describes diagnosis identification in the sample, not error-free reports across the entire dataset. The generated reports still contained mistakes, which the radiologists reviewing and editing them found easy to catch, according to the university’s account.

A separate blinded evaluation tested whether experts could distinguish AI-written reports from those written by radiologists. Panelists could identify the unedited AI output. After radiologists edited those drafts, however, the experts usually could not tell them apart from reports written by radiologists. The comparison therefore concerns both raw AI output and a human-edited version—not just the machine working alone.

The goal is not to replace the physician. We're not trying to build an AI radiologist.

Harrison Bai, senior author and visiting associate professor of radiology

The emergency-care promise needs a clinical test

The team’s intended workflow keeps a trained radiologist in charge of evaluating the preliminary report. Bai argues that letting AI prepare the draft could save time. That is a proposed benefit of the collaboration, rather than a reason to remove the review step that was part of the study’s report evaluation.

The researchers also want the system to help identify patients who may need urgent surgery and move them toward treatment sooner. In the University of Colorado account carried by News-Medical, that ambition sits alongside an explicit next step: evaluating performance in real-world emergency care before wider adoption.

That evaluation must examine the physician–AI partnership, not only the agent’s answers. The researchers identify diagnostic accuracy, efficiency and patient well-being as outcomes requiring further study. Producing a useful draft and helping a patient receive better care are related goals, but this study leaves the latter question open.