https://doi.org/10.1007/s00392-025-02625-4
1Justus-Liebig-Universität Giessen Medizinische Klinik I, Kardiologie Bad Nauheim, Deutschland; 2Kerckhoff Klinik GmbH Abteilung für Kardiologie Bad Nauheim, Deutschland; 3Universitätsklinikum Gießen und Marburg GmbH Medizinische Klinik I - Kardiologie und Angiologie Gießen, Deutschland
Background and Aims: Coronary angiography (CA), an essential diagnostic and therapeutic tool in cardiology, relies on X-ray sequences. Optimal visual results are achieved through high-frame-rate video sequences with the downside of higher radiation dose for both patients and examiners. An established artificial intelligence (AI) method in video processing is optical flow estimation for frame interpolation. Goal of this project is to to transfer this AI method to CA sequences aiming at reduction of X-ray exposure while maintaining smooth video quality by frame interpolation.
Methods and Results: Data of a registry cohort that enrolled patients scheduled for CA due to a suspected chronic coronary syndrome was used. 100 patients were selected in whom 365 sequences with a framerate of 10fps were available and used as ground-truth. Based on theses, sequences with 5fps were generated by removing each second frame. These low-fps sequences were than interpolated to 10fps using published general purpose real-time intermediate flow estimation (RIFE) AI based framework. Two common computer vision metrics—PSNR and SSIM—are applied to objectively assess the quality of the interpolated frames. Currently, a survey is being conducted with cardiologists evaluating the smoothness and accuracy of the interpolated videos in comparison with the ground-truth sequences. The interpolated 10fps videos are based on only 5 x-ray images per second, therefore a relevant radiation dose reduction is achievable. Preliminary data shows a high quality of interpolated CA sequences both in the technical and human evaluation.
Conclusions: Enhancing the framerate of CA sequences via AI based frame interpolation is feasible and provides sufficient quality thereby showing the potential for radiation dose reduction without loosing image quality. However, further research is needed to detect or avoid undesired interpolation and elemental artifacts as well as to provide a potential real-time use.