Using AI Could Help Patients Understand Radiology Reports

Patients who read radiology reports before speaking with their physicians can struggle to interpret the medical language and may become unnecessarily alarmed.

A study led by researchers at the University of California, San Francisco (UCSF) suggests summaries of radiology reports created using AI could increase patients’ understanding and reduce their worries.

The 21st Century Cures Act, enacted in 2016, requires that patients have timely access to their health information, including radiology reports. While that access promotes transparency, it can also create challenges when patients encounter concerning findings, especially if they consult the reports before discussing them with the ordering physician.

Because radiology reports are written for clinicians, they often contain medical terms that may confuse patients and can result in unnecessary worry. “The risk is that patients won’t understand what is normal and what is abnormal,” said Juan Serna, MD, the study’s lead author.

“There’s more emphasis on democratization of patient information,” said corresponding author, Jae Ho Sohn, MD, MS. “It’s really good in theory, but the law also created challenges.”

Dr. Sohn said colleagues in the UCSF Department of Radiology and Biomedical Imaging and patients themselves have described the distress caused by reading the reports.

In addition, Dr. Sohn’s primary care colleagues are also fielding more phone calls from patients about the reports. “The primary care doctors are super helpful for discussing reports, but they can’t micromanage every scan for every patient,” Dr. Serna observed.

Listen as Dr. Serna discusses his research.

AI Reports Lessened Concerns, Increased Clarity

To test whether AI could reduce confusion, Drs. Serna and Sohn and colleagues used three deidentified reports from lung cancer CT screenings taken from the UCSF radiology database, selected to reflect a diverse range of pathologies.

They employed a secure version of ChatGPT, the popular AI chatbot, to create simplified versions of each of the reports. A radiologist reviewed the AI versions for accuracy.

The researchers chose lung cancer screening reports because they have a high potential to cause anxiety and often contain findings that can be confusing. “The reports covered a range of findings, like ‘negative for cancer,’ ‘the third one was highly suspicious for cancer’ ‘second one was mixed picture,’” Dr. Serna said.

The researchers tested the reports with participants recruited from Prolific, an online database of study volunteers. The participants were screened to create a pool that was demographically nearly representative of the U.S. population.

For each of the three findings, the participants read the original report, then the AI version. They answered survey questions after they were done with each report. The surveys were conducted over five days in July 2025.

The researchers compared the answers of 1,815 respondents and found that the AI-generated summaries improved participants’ comprehension and decreased their anxiety.

The summaries also increased participants’ hypothetical willingness to wait for a scheduled follow-up appointment to go over the results, rather than call their doctor immediately the same day.