Abstract 4143663: Quantifying the Coronary Artery Slow Flow Phenomenon: Insight from Dynamic Angiography and Deep Learning

Background: The coronary slow-flow phenomenon is frequently seen in patients with coronary stenosis, coronary endothelium dysfunction, microvascular dysfunction, etc. It is diagnosed as TIMI-2 flow grade (requiring ≥ 3 cardiac cycles to opacify the vessel) or a corrected TIMI frame count (TFC) greater than 27 frames. However, the TFC has variability due to its dependence on injection force. In cases without autonomic injection, how could to quantify frame count with minimal bias? Methods: All patients underwent the angiogram included, and the exclusive criteria were patients with heart failure, cardiomyopathy, congenital heart disease, valvular disease, and multivessel disease. The angiographic recording began with the injection of contrast into the coronary artery and continued with the white-colored blood moving until all the contrast had been washed out from the arteries. The review of coronary was 15 frames per second and focused on the white-colored blood flow against a black contrast background. The arterial phase (AP) calculation started when the blood began moving and ended when all the contrast was washed out of the distal arterial vasculature. The artificial intelligence program was built based on U-net and Densenet-121. At first, a segmentation model and a Convolutional neural network (CNN) model are responsible for detecting the starting frame of AP with a full index of contrast. In the end, the CNN model will detect the ending frame of AP. Results: 151 patients met the inclusion criteria, and 76.8% were males. Subgroup analysis of patients with patent coronary arteries showed that the AP of the group with slow flow at the distal segment was 32.3 ± 4.4 frames. There was a statistically significant difference compared to the no-slow-flow group (p<0.001). The AP of patients with patent coronary artery was 25.8 ± 4.4 frames, compared with the AP of the group with a significant lesion, which was 32.2 ± 4.0 frames (p<0.001). The cut-off value of the AP frame count in diagnosing coronary slow flow was 27.5 frames, with a sensitivity of 91% and a specificity of 67%. Moreover, the Deep Learning model for calculating AP frames was successfully trained, achieving a root mean square error of 5.1 frames. Conclusion: In addition to TIMI frame count, AP frame count analysis is feasible for cath labs without autonomic injection. The preliminary results suggest AP is a potential marker for diagnosing the slow flow phenomenon.