Detection of lipid and fibrous plaques in optical coherence tomography images using deep learning

Abstract

The study goal was to evaluate the possibility of automating of optic coherent (OCT) images analysis and to develop a model for precise differentiation of nature of intracoronary atherosclerotic plaques using deep neural networks. At the first stage of the study, special software was developed that allows the experts to mark 16 different types of pathologies in the intracoronary images obtained from OCT. The following pipeline was used for image preprocessing: 1) image smoothing using a Gaussian filter (noise reduction), 2) image binarization using a threshold of values determined according to the Bradley-Roth adaptive (local) binarization method; 3) interpolation of missing values; 4) vascular wall determining; 5) transformation of polar coordinates into Cartesian ones. 534 images were collected for analysis, obtained as a result of OCT of the main 3 coronary vessels of the heart. Two models (convolutional neural networks with the AlexNet and ResNet-18 architecture) were trained to detect the pathologies. The results demonstrate high accuracy of the both models: plaques detection 0.72-1.00; lipid plaque - 0.875-0.929 and fibrous differentiation - 0.893-0.929.