Abstract 12868: Real-Time Explainable Detection of Stenosis in Coronary Angiography

Abstract

Background: Coronary angiography is the gold standard for evaluating coronary artery disease and determining the appropriate treatment strategy for a patient. The aim of this study is to use a machine learning (ML)-based object detection algorithm to identify images with clinically significant stenoses and provide explainable output by localizing the area of stenosis within each image. Methods: We utilize the YOLOv5 object detection algorithm to detect clinically significant stenoses by treating narrowing of the vessel as the object to be located in an image. Models are trained utilizing a 5-fold cross-validation strategy on a dataset consisting of the middle frame from 756 randomly selected cine clips from an interventional cardiology suite. Maximum object confidence on the validation dataset is used for downstream whole image classification. Results: Across 5 folds, mean bounding box precision was 0.342 ± 0.074 and recall was 0.305 ± 0.021 at an intersection over union threshold of 0.6. The image classification model achieved a mean and standard deviation area under the receiver operating characteristic curve of 0.863 ± 0.042 with a mean accuracy of 0.809 ± 0.041 at the optimal confidence cutoff of 0.283 ± 0.241. Additionally, trained model predictions occurred at a rate of 32.8 ± 3.6 frames per second, allowing for live model inference and provider support at a standard angiographic frame rate of 15 frames per second. Conclusions: The proposed object detection ML approach is a time-efficient and data-efficient method of diagnosing significant coronary artery stenosis. These results combine state-of-the-art performance in detecting stenosis in angiographic images with improved explainability over single-step classification with a deep neural network.