Imaging panel discusses pace, direction of practical AI applications

Panelists reviewed the state of artificial intelligence (AI) in radiology and prospects for its future application in respiratory health during the CHEST 2025 session Radiology AI: Moving Forward With Momentum on Wednesday, October 22, in Chicago.

AI in chest radiography

Chest radiography is the most performed medical image test around the world, with studies estimating that 2 billion to 4 billion chest X-rays are performed each year, said Bruno Hochhegger, MD, PhD, Clinical Professor of Radiology and Vice Chair of Research for the Department of Radiology at the University of Florida.

“Certainly, there are not enough doctors to see all these X-rays,” he said, adding that the problem becomes even more pronounced with time demands and staff shortages.

Dr. Hochhegger said traditional image diagnosis is challenged by subtle findings, perceptual errors, logistical hurdles for whole-scale implementation, and by lack of clarity about legal liability. Quality control is one of the most promising AI applications in chest medicine, he said.

“We have a lot of data about using AI as a peer review or a second reader showing that this could increase the quality of reports for radiologists, for pulmonologists, and for ER physicians,” Dr. Hochhegger said.

Though there are barriers and a need for trials about patient outcomes, he said the use of AI in chest medicine is “real, validated, and expanding.”

AI in ILD

There is a rapidly narrowing difference between AI and human capabilities for medical imaging applications, including the identification of nodules for interstitial lung disease (ILD), said Kevin K. Brown, MD, FCCP, Professor and Chair of the Department Medicine at National Jewish Health in Denver.

Ideally, he said, the next step is to apply AI to help identify biomarkers for predisposition, diagnosis, prognosis, and therapeutic response. Procedures used in identifying these biomarkers would need to lead to affordable and practical steps to improve patient care. Dr. Brown, however, highlighted a “cautionary tale” from accelerated approval pathways in oncology, where most therapies that had been approved based on a surrogate imaging marker did not demonstrate an overall survival benefit in subsequent studies.

Dr. Brown also cited a report from a 2023 symposium led by the US Food and Drug Administration on meaningful end points in idiopathic pulmonary fibrosis that acknowledged the utility of imaging biomarkers in screening, diagnosis, and—potentially—enrichment of clinical trials. However, the report highlighted the importance of defining risk and tolerance for uncertainty when using any radiological biomarker and noted that imaging biomarkers have not been demonstrated to be superior or noninferior to other endpoints.

Looking to the future, Dr. Brown said, the use of AI for chest imaging biomarkers faces a confirmation pathway for clinical trial and clinical use that is necessarily complex.

“But don’t get me wrong—there is enormous promise here,” Dr. Brown said. “It is cutting edge. We are not going to stop it. We are going to figure out how to incorporate it in a way that is most safe and most beneficial for patients.”

AI in lung cancer

Amit Gupta, MD, MRMD, CIIP, noted a growing and evolving application of AI imagery in early-stage lung cancer detection and shared thoughts about structural approaches for the optimal integration of these applications into health care settings. 

Lung cancer remains the no. 1 cause of cancer deaths worldwide—and several studies have documented the effectiveness and the underused potential of early low-dose screening, said Dr. Gupta, Division Chief of Cardiothoracic Imaging and Modality Director of Diagnostic Radiography at University Hospital Cleveland Medical Center and Associate Professor of Radiology, Medicine, and Biomedical Engineering at Case Western University.

As the performance capabilities of AI imagery improve, the technology can make a crucial difference in this area. But to be most effective, AI’s known strengths and limitations should be seamlessly integrated into the end-to-end picture archiving and communication system workflow of an institution, with plans for education, feedback, and monitoring built into the process. And once an AI imaging system for detection and analysis is adopted and integrated, it must be monitored and regularly revised.

As an example of how an AI system should respond to the needs and system of those who use it, Dr. Gupta detailed how AI imagery reports were adapted to include a coronary artery calcium calculation. This calculation was visible to radiologists at the initial level but could be toggled off for readings at subsequent levels, if those findings were not significant.

“No tool is perfect. No human is perfect,” Dr. Gupta said. “So why not use what we have in unison and apply all this practical knowledge to help our clinicians and our patients?”