AI cuts coronary angiography radiation by more than half

A research team led by Prof. Kang Si-hyeok of the cardiology department at Seoul National University Bundang Hospital — with researchers Kwon Hwi and Park Se-young as lead authors — has developed an AI-based video interpolation model called Angio-FILM. The model produces smooth coronary angiography footage even when captured at low frame rates that cut radiation exposure by more than half.

Coronary angiography is a diagnostic procedure in which a contrast agent is injected and X-ray video is used to examine the shape and blood flow of the heart's vessels.

The procedure is used to observe blood vessels in detail during the diagnosis and treatment of coronary artery diseases such as myocardial infarction. To keep pace with the rapid beating of the heart and coronary arteries, the imaging system captures 10 to 15 frames per second — but the higher the frame rate, the greater the radiation exposure for both patients and medical staff.

Lowering the frame rate to reduce radiation creates longer gaps between captured images, which can cause vessel motion to appear choppy or jittery. Concerns about image quality are precisely why low-frame-rate acquisition has not been widely adopted in clinical settings. Patients are directly exposed to radiation during the procedure, and medical staff must endure repeated daily exposure while wearing heavy protective gear to perform precise interventions.

Angio-FILM addresses this by using AI to generate the intermediate frames that would exist between captured images, cutting the radiation dose while maintaining image quality at conventional levels.

The system captures footage at 7.5 frames per second — half the current standard — then uses AI interpolation to fill in the missing frames, restoring the equivalent of 15-frames-per-second image quality. The research team estimates this approach can reduce radiation exposure by more than half.

"Because this footage is used in precise interventional procedures, the model was designed to stably reproduce the fast and nonlinear motion of the heart and coronary arteries," the research team said. Rather than simply averaging adjacent frames, Angio-FILM applies a technique called Latent Flow Matching, which separates spatial and temporal analysis algorithms and computes motion paths using only the key elements of the image — an approach the team said improves stability.

In a Turing test, 30 specialist physicians were asked to distinguish between 600 original angiography clips and AI-interpolated versions. Even when explicitly told that AI had been applied, the physicians identified the AI-generated footage at a rate no better than random chance — a 50 percent hit rate with no statistically significant difference.

The result suggests the reconstructed images are precise enough that even trained specialists looking for AI artifacts cannot reliably detect them. The margin of error in coronary lumen diameter between original and AI-generated images was just 0.18 millimeters, allaying concerns about anatomical distortion.

The study is significant for offering a practical solution that cuts radiation exposure from coronary angiography by more than half while still producing images suitable for clinical use.

Coronary artery disease is a life-threatening condition and one of the core areas of so-called essential medicine, which has faced mounting strain in recent years. Protecting both patients and medical staff from radiation exposure has been a major challenge, and AI has opened a path forward.

"Physical equipment improvements aimed at reducing radiation in coronary angiography have already reached their limits," Kang said. "Having established Angio-FILM's clinical reliability through this research, we believe its introduction into clinical settings could dramatically reduce radiation exposure for both patients and medical staff."

The findings were published in the latest issue of npj Digital Medicine, a sister journal of Nature.