AI and Chest X-Rays: A Strong Co-Pilot, but Is it Ready to Fly Solo?
An RSNA 2025 session explored whether AI should be able to autonomously interpret chest radiographs
As AI continues its rapid climb in medical imaging, a spirited debate at RSNA 2025 asked: Is AI ready to take the pilot’s seat in interpreting chest X-rays (CXR)? Or, does it still need a human co-pilot to ensure patient safety and diagnostic accuracy?
According to an informal poll conducted by the RSNA Daily Bulletin, 68.4% of respondents agreed with Warren Gefter, MD, a radiologist on the emeritus faculty at Penn Medicine in Philadelphia, who said that current AI models are not sufficiently accurate, reliable, trustworthy or comprehensive enough for totally autonomous AI (CXR) interpretation.
However, 31.6% of those who responded sided with Saurabh Jha, MD, associate professor of radiology at the Hospital of the University of Pennsylvania, in Philadelphia. Dr Jha argued that because CXR are no longer used for diagnosis, interpreting or reporting them no longer requires a medical degree or radiology training. “This is something that AI is more than capable of doing autonomously,” he said.
So, who is right?
To try to find out, we asked Eun Kyoung (Amy) Hong, MD, PhD, a thoracic radiologist at Stanford Medicine in California. “Although I don’t think I’m qualified to say whether one side is right and the other wrong, what I can do is provide my perspective based on years of working with and studying AI’s use in chest X-ray interpretation,” she said.
The Limits of Generative AI
Before one can debate autonomy, Dr. Hong said we first need to understand what AI can—and cannot—do.
“On one hand, we have conventional vision-based AI models, which can support the classification, detection and segmentation of specific abnormalities but require a radiologist to interpret the outputs and create the report,” Dr. Hong explained. “On the other hand, there are the generative AI models that are trained on paired image and data sets and are used to produce preliminary reports for us to review before signing them.”
While Dr. Hong agrees that generative AI models can do far more than traditional vision systems in CXR interpretation, she warned that this advantage comes with a myriad of problems.
One of those problems is hallucinations. “I define a hallucination as the mentioning of something that was not included in the input given to the model, such as referencing prior imaging that never happened,” Dr. Hong said.
Based on several peer reviewed studies, Dr. Hong estimates that hallucinations or unsupported statements appear in 15–20% of all generated reports.
While this is a significant problem, and one that tilts the needle towards AI not being ready to autonomously interpret radiographs, Dr. Hong noted it is likely a temporary issue. “The technology is advancing very fast, so I believe that the problem of hallucinations will be solved very soon,” she said.
Acceptability Is Key
Dr. Hong also acknowledged that generative models can improve efficiency.
She cited a large prospective trial of nearly 12,000 CXRs that were read by 22 radiologists and an AI assistant that reduced interpretation time from about 189 seconds to 160 seconds.
In the trial, AI helped facilitate an approximately 15% gain in efficiency. But this trial was limited to the use of AI in assisting radiologists in interpreting reports. It did not address whether AI could interpret the reports autonomously.
On that point, Dr. Hong highlighted a study where over 1,500 AI generated CXR reports were read by thoracic radiologists, with 64% of the reports accepted without modification. She observed that this finding shows radiologists are still revising many of the drafts, a practice which does little to advance efficiency.
“I believe that acceptability is key in terms of an AI generated report being helpful and able to increase efficiency. Reviewing and revising a report can be just as burdensome—if not more burdensome—than creating the report from scratch,” Dr. Hong said.
A Huge Gap
While acceptability is important, it’s not the only issue that needs to be addressed before AI is allowed to autonomously interpret CXRs. There's also the issue of model variability.
Dr. Hong recently compared four different generative AI models. These included two agent-based models, one domain specific radiological report generation model and a generalist model.
Agent-based AI focuses on autonomous decision-making and task execution over time, while domain-based AI focuses on deep expertise and accuracy within a specific field. Generalist models are designed to perform competently across many domains and tasks without specialization.
Dr. Hong found that acceptability ranged from 28% to 67% and hallucination from 5% to 55%, which she noted as being a huge gap.
“Which model do we choose to embed in our practice to make chest X-ray reads autonomous? That's a very big discussion and one that requires rigorous benchmarking,” she said.
Other issues highlighted by Dr. Hong include the risk of automation bias and the lack of legal framework for assigning responsibility for missed findings by AI.
A Weak Soloist
Based on her extensive research, Dr. Hong believes that, when it comes to interpreting CXRs, AI makes for a strong co-pilot, but it’s not yet ready to fly solo.
“AI can help us draft reports, standardize terminology and reduce operational burdens, but it’s still a weak soloist,” she said.
To back her opinion, she pointed to the fact that generative AI models cannot take legal responsibility for their interpretations, and she emphasized that the models still struggle with certain rare and subtle findings, which they fill in with hallucinations.
“Autonomy requires zero hallucination, consistent findings and a clear legal and regulatory framework, which we just don’t have yet,” Dr. Hong concluded. “Until we do, AI can be used to help draft and interpret reports; it’s just not ready to sign them.”
For More Information
Read the related Daily Bulletin article, “Using AI to Autonomously Interpret Chest X-Rays: Too Soon or Not Soon Enough?”
Read previous RSNA News articles on the use of AI with chest X-ray: