When a patient arrives in the emergency department with severe abdominal pain, every minute counts. Researchers at the University of Colorado Anschutz School of Medicine are investigating ways to streamline the work of radiologists and make it more efficient. They also aim to quickly identify patients who may need urgent surgery and speed up the path to treatment.
The scientists have uncovered a new way to leverage artificial intelligence (AI) to interpret right upper quadrant (RUQ) ultrasounds, which visualize organs in the right upper side of the abdomen such as the liver, the gallbladder, the right kidney, and part of the pancreas.
In a recent study in Nature Communications, researchers created a vision-language AI agent, which blends visual ultrasound data with written clinical data from the lab, that can accurately read RUQ ultrasounds, diagnose common health conditions affecting these organs, and suggest logical next steps.
Harrison Bai, MD, MS , a visiting associate professor in the Department of Radiology and senior author of the paper, emphasizes that this AI agent is designed to work alongside a trained radiologist, generating reports for the radiologist to read and evaluate. It's a tool that can help streamline radiologists' daily work.
"The goal is not to replace the physician. We're not trying to build an AI radiologist," Bai says. "But if our AI agent can write up preliminary reports, it saves radiologists time."
And time is of the essence, especially in the case of a medical emergency: Bai and his team's new AI tool can also rapidly identify patients who may need emergency surgery so they can get to the operating table as quickly as possible.
Why integrate AI into RUQ ultrasound interpretation?
RUQ ultrasounds are often doctors' top choice for diagnosing issues involving the liver or biliary system. Doctors often order them when patients are experiencing pain in the right upper abdomen or have had certain types of abnormal lab results, such as elevated liver enzymes or bilirubin.
These ultrasounds are painless and non-invasive, and they can help diagnose a range of conditions such as gallstones, fatty liver, liver cirrhosis, hepatitis, kidney stones, or pancreatitis. RUQ ultrasounds can even potentially find cysts or tumors in these organs.
Many fields of medicine are facing shortages of specialists, but radiologists are facing extra pressure due to an aging U.S. population and a rising demand for medical imaging. Radiologists are working harder than ever as patient loads increase and doctors are asked to accomplish more with fewer resources. As of 2022 , radiologists were leaving the field at twice the rate they were in 2014.
"We're trying to address the increasing workload. It takes time to write reports on imaging results. There aren't enough people or radiologists to do the work, and there's decreasing reimbursement by Medicare insurance," Bai says. "If you tell radiologists that an AI tool can save them time, suddenly, people are much more likely to use the technology."
Interpreting a RUQ ultrasound can also be challenging. The people who interpret them often have varying levels of skill and experience, and the ultrasounds themselves can have significant visual noise and artifacts that less-trained interpreters may not recognize.
According to Bai and his fellow researchers, that means the diagnostic accuracy of an RUQ ultrasound often depends on who's reading it. Incorporating an AI vision-language model (VLM) could help standardize how these ultrasounds are interpreted, improve accuracy, and make results more consistent across different ultrasound interpreters.
The VLM the researchers created can also generate reports quickly, which is especially crucial when patients are experiencing a medical emergency. For example, patients with acute gallbladder inflammation (cholecystitis) need emergency surgery right away. These patients can receive their diagnosis and get the surgery they need faster with the help of the VLM.
AI: A valuable collaborator, not a replacement, for radiologists
In the paper, the researchers discuss how the VLM should operate as a "robust, clinically integrated assistant" that can integrate different types of information and help with clinical decision-making, but without replacing the expertise that a radiologist provides.
The team designed the VLM to read RUQ ultrasounds, scan them for 16 common diagnoses, estimate each patient's prognosis, and auto-generate reports. Researchers tested the VLM on a dataset of roughly 11,000 RUQ ultrasounds from Stanford University, University of Colorado, and Johns Hopkins University and then manually reviewed 100 of the AI-generated ultrasound reports for accuracy.
In that sample of reports, the VLM did remarkably well in some respects: It identified the correct diagnoses in each ultrasound 99.6% of the time. The VLM made some mistakes in the overall reports, but these mistakes were easy to catch for the radiologists who reviewed and edited the reports.
The researchers also asked a panel of experts to participate in a blinded evaluation to see if they could distinguish VLM-written reports from reports that radiologists wrote themselves. While the experts could pick out unedited, VLM-written reports, most of the time, they could not tell the radiologist-edited VLM reports apart from the reports written by radiologists.
Before this technology can be widely adopted, researchers need to evaluate its performance in real-world emergency care. The researchers say it will also be critical to study how physician-AI collaboration affects diagnostic accuracy, efficiency, and patient well-being.
The study, however, suggests AI has a real place in supporting radiologists without eliminating the need for their expertise. Bai says he doesn't expect AI to replace the need for human supervision in medicine, and he knows it will take time for doctors to adapt to using the technology, but he thinks it could change the course of medicine for the better.
"I can see this technology playing a huge role in how medicine is practiced," he says. "It will involve a lot of change and learning about new things. But I'm super excited. I want to see these developments not just happening in my lab, but also incorporated into daily practice to help patients and doctors."
