Neural network analysis of coronary artery stenoses: assessment of the accuracy and speed of promising architectures

Abstract

Significant interest in the field of application of machine learning for the analysis of medical images stimulates the search for promising algorithms for solving routine diagnostic problems in cardiology. In relation to cardiovascular diseases, such a procedure is coronary angiography, which assesses the state of the vascular network and the presence of stenotic areas. This paper demonstrates an example of using modern models of neural networks: SSD MobileNet V2, SSD ResNet-50, Faster-RCNN Inception ResNet for localizing a single-vessel coronary artery lesion on a set of clinical data (3200 images). It is shown that the Faster-RCNN Inception ResNet V2 model was the most accurate in terms of the chosen metric mAP[0.5:0.95], reaching 0.9434 and 0.95 for the validation and test sets, respectively. However, the data processing speed was 0.363 seconds per frame, which corresponds to a speed of 2.8 frames/sec, which does not correspond to the speed of coronary angiography (15 frames/sec). Neural networks with a more “simple” architecture demonstrated an unsatisfactory quality of stenosis localization, expressed in a low characteristic mAP[0.5:0.95]. The results of this study demonstrate a key problem in the application of machine learning algorithms on graphic data – high accuracy, which may be acceptable for medical diagnostic procedures, is “decompensated” by long-term image analysis, as a result, the use of unmodified neural network architectures does not provide real-time data processing.