Using AI to Autonomously Interpret Chest X-Rays: Too Soon or Not Soon Enough?

Should AI be used to autonomously interpret and report chest radiographs? That was the topic of a lively debate at a Thursday RSNA session.

AI is changing chest radiography (CXR) as we know it. In some cases, it can already detect a variety of findings with performance comparable to—or even exceeding—that of human radiologists. 

But is AI ready to be a chest X-ray jedi? According to Warren Gefter, MD, a radiologist on the emeritus faculty at Penn Medicine in Philadelphia, the answer is ‘not so fast’. 

 “Totally autonomous AI, where algorithms interpret all chest X-rays without a radiologist’s input, is a pipe dream for the foreseeable future,” he said. 

According to Dr. Gefter, despite remarkable progress and potential, current AI models are not sufficiently accurate, reliable, trustworthy or comprehensive enough for totally autonomous AI CXR interpretation. 

“Numerous hurdles need to be overcome—and new AI approaches introduced—before we allow AI to interpret CXRs without human input,” he explained. “AI does not yet have the reasoning capability or the contextual understanding of images required for the diagnostic interpretation of CXR findings, particularly for more complex cases.” 

The question we should be asking, argued Dr. Gefter, is ‘how can radiologists and AI work together to create a system greater than the sum of its parts?’

“A radiologist-AI collaborative approach to CXR reporting and workflow is more feasible and, when optimized, can improve efficiency as well as the accuracy, trustworthiness and information content of CXR reports,” Dr. Gefter said. 

Recognizing that radiologists are facing an unsustainable, increasing workload, Dr. Gefter acknowledged the need for improved efficiency. But instead of handing over the reins to AI and sacrificing the accuracy and quality of CXR interpretations, he said we can still make efficiency gains by, for example, having the AI draft reports and assisting in integrating clinical and prior imaging data. “It’s the best of both worlds,” he added. 

Dr. Gefter concluded by noting that because chest radiographs are the most frequently performed diagnostic X-ray examination worldwide, there is a strong motivation to offload their reporting to autonomous AI. “It is critically important that AI be deployed cautiously and carefully and in a manner that not only improves radiologists’ efficiency but also ensures that it is safe and beneficial to patients,” he said.

Time to Give the Keys to AI 

On the other side of the table is Saurabh Jha, MD, associated professor of radiology at the Hospital of the University of Pennsylvania, who argued that it’s exactly because chest radiographs are so frequently performed that they should be fully automated.

According to Dr. Jha, the chest X-ray is so ubiquitous that it no longer has any inferential value, that it’s become a data point used to manage disease. “Formerly a diagnostic tool, chest X-rays are now used to verify procedural success—not even nasogastric tubes are placed without radiographic guidance,” he explained.

Because chest X-rays are no longer used for diagnosing, Dr. Jha argued that 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.

With this accuracy in mind, Dr. Jha believes that requiring AI-human collaboration is unnecessarily redundant. In fact, some studies suggest that when a chest radiologist uses AI their performance actually decreases. “The whole point of AI is to increase efficiency,” he said. “If we require that a radiologist check every chest X-ray, then what’s the point of using AI in first place?”

This is why Dr. Jha argues that it’s time to give the keys to AI. 

“Radiologists spend too much time on low-complexity cases and too little on high-complexity ones,” he concluded. “Using AI to autonomously interpret and report chest X-rays is one area where we can help tilt the scale in the right direction and we would be remiss not to.”

Access the presentation, “Should AI Autonomously Interpret Radiographs? A Debate,” (R3-CNPM14) on demand at RSNA.org/MeetingCentral.