Abstract
BACKGROUND: Coronary artery disease (CAD) is a leading cause of global mortality, primarily due to the accumulation of atheromatous plaques in coronary arteries. The current diagnostic standards include X-ray coronary angiography, which evaluates morphological features of the coronary arteries. The primary goal of this modality is to quantify stenosis severity, but variability and complexities in the resultant images make consistent interpretation challenging. This study developed a deep learning-based approach for segmenting and grading vessel stenosis via X-ray coronary angiography. METHODS: Based on 383 angiographic images from 168 patients, we developed a dual-output deep convolutional neural network (CNN) to automatically diagnose stenosis. For clinical relevance, we manually annotated stenosis severity into five distinct levels: nonobstructive lesion (1-49%), intermediate lesion (50-70%), severe lesion (71-95%), sub-total occlusion (96-99%), and total occlusion (100%). RESULTS: We built a coronary stenosis segmentation and grading method based on X-ray coronary angiography. The model achieved an average intersection over union (IoU) of 0.92, a Dice score of 0.95, a precision of 0.93, and a sensitivity of 0.96. CONCLUSIONS: We introduce an image-based vascular analysis method that localizes and grades stenosis in X-ray coronary angiography. This method can automatically identify clinically critical grades, especially within the 71-100% range. Deep learning methods can potentially facilitate the diagnosis of CAD and patient-centric treatment planning.