Convolutional Neural Network for Atherosclerotic Plaque Multiclass Semantic Image Segmentation in Transverse Ultrasound Images of Carotid Artery

Arterial stenosis is one of the most common diseases and if it is not discovered in time and adequately treated, it may have critical consequences, such as a debilitating stroke and even death. This is the reason why early detection is a number one priority. This disease occurs as a result of plaque deposition within the coronary vessel. The process of manually annotating plaque components is both resource and time consuming, therefore, an automatic and accurate segmentation tool is necessary. The goal of this research is to create a model that sufficiently identifies and segments atherosclerotic plaque components such as fibrous and calcified tissue and lipid core, by using Convolutional Neural Network (CNN) on transverse ultrasound imaging data of carotid artery. U-net model was trained with dataset of 60 ultrasound samples, collected and annotated by medical experts during TAXINOMISIS project, and achieved 96.94% and 57.38% Jaccard similarity coefficient (JSC) for segmentation of background and fibrous classes, respectively. On the contrary, model had difficulties with segmentation of lipid and calcified plaque components due to dataset being imbalanced and small, which is shown with respective JSC values of 19.05% and 32.68%. Future research will focus on expanding current dataset with additional annotated ultrasound samples, with the goal of improving segmentation of lipid and calcified plaque components.