AI-Driven Analysis of Coronary Artery Stenosis on Cardiac CTA

Radiologists and cardiologists have struggled to keep pace with the rapidly increasing volume of cardiac CTA studies worldwide. AI offers a potential tool to alleviate the workload by automatically checking the images for obstructions and quantifying the amount of calcium in coronary arteries. 

To learn more about this potential, researchers studied a prototype AI-based tool that automatically analyzes coronary artery stenosis on cardiac CTA. 

They applied the AI tool to all the cardiac CTA studies performed between 2017 and 2021 at a single academic center, excluding non-diagnostic tests, coronary anomalies and patients who had stents and post-bypass grafts. The final study group included 1,041 cardiac CT studies from 1,032 patients. CAD-RADS categories extracted from structured clinical radiology reports served as the reference standard. 

The AI achieved excellent diagnostic accuracy, with an area under the receiver operating characteristic curve of 0.90 for detecting maximum stenosis of both 50% or more and 70% or more. 

“I think the main message here is that the AI is getting good at this,” said study lead author Fernando Kay, MD, PhD, from the University of Texas Southwestern Medical Center in Dallas. “An area under the curve of 0.90 is really promising.”

The algorithm was particularly effective at predicting which patients did not have significant stenosis.

“We found that about 98% of the time, people who had a negative study on AI did not have any major obstruction,” Dr. Kay said. “That’s very close to what an experienced human observer would achieve.” 

In contrast, the algorithm’s positive predictive value was low. Only 39% of patients identified as having obstructive disease by the algorithm actually had an obstruction.

“This tool will be a great help to radiologists, but it’s making some mistakes in the positive predictors, so there are some opportunities for improvement down the road,” Dr. Kay said.

The algorithm’s ability to rule out obstructions makes it a potentially useful tool in emergency departments, Dr. Kay said, especially during hours when there are no cardiac imagers available. The tool could provide an initial assessment, possibly sparing patients from longer stays in the hospital.

“We envision this as a tool that would help triage patients,” Dr. Kay said. “You could quickly do the cardiac CTA on the patient, run this algorithm and if the result is normal, then the patient could be discharged from the hospital, avoiding additional costs and downstream testing.” 

Dr. Kay said that future studies are required to investigate how AI-preliminary read strategies impact patient management, with a focus on safety and turnaround time.

Access the presentation, “Diagnostic Performance of AI-Driven Analysis of Coronary Artery Stenosis in Cardiac CTA,” (S5-SSCA02-4) on demand at RSNA.org/MeetingCentral