Feasibility of the Anchor-Free Deep Learning Method in Coronary Stenosis Automatic Detection

Abstract

Background. Coronary artery disease (CAD) is a type of cardiovascular disease which is one of the leading causes of death around the world. The presence of coronary stenosis is considered a pivotal indicator in the diagnosis of various CADs. The main purpose of this paper was to investigate the feasibility of an anchor-free deep learning (DL) method, fully convolutional one-stage object detection (FCOS), in coronary artery stenosis automatic detection. Methods. First, 2786 invasive coronary angiography (ICA) images from 130 patients were randomly divided into training, validation, and testing datasets using the 10-fold cross-validation approach. Then, FCOS was compared with other three widely used anchor-based DL models: single shot multibox detector (SSD), faster region-based convolutional network (Faster R-CNN), and you only look once (YOLOv3), in terms of precision, recall, F1 score, average precision (AP), and average recall (AR). Finally, the performances of different models in the detection of stenosis were compared in either single or multiple lesion scenarios using statistical tests. Results. FCOS achieved significantly superior precision (96.14% ± 0.53%), recall (94.36% ± 0.79%), F1 score (95.22% ± 0.56%), AP_0.50 (93.36% ± 0.93%), AR_0.50:0.95 (64.73% ± 1.46%), AP_small (55.04 ± 0.96%), AP_medium (59.97 ± 1.13%), and AP_large (68.09 ± 5.18%) compared to Faster R-CNN and YOLOv3. Moreover, FCOS demonstrated significantly higher AR_0.50:0.95 and AP_small compared to SSD. Regardless of the presence of single or multiple coronary stenoses in ICA images, FCOS also outperformed Faster R-CNN and YOLOv3. Furthermore, it showed significantly higher AR_0.50:0.95 compared to SSD when in the multiple stenosis scenario. Conclusions. It is feasible to use the anchor-free DL model FCOS in detecting coronary stenosis based on ICA images.