Atherosclerotic plaque deposition within the coronary vessel wall leads to arterial stenosis and if not adequately treated, it may potentially have deteriorating consequences, such as a debilitating stroke, thus making early detection of the most importance. The manual plaque components annotation process is both time and resource consuming, therefore, an automatic and accurate segmentation tool is necessary. The main aim of this paper is to present the model for identification and segmentation of the atherosclerotic plaque components such as lipid core, fibrous and calcified tissue, by using Convolutional Neural Network on patch-based segments of ultrasound images. There was some research done on the topic of plaque components segmentation, but not in ultrasound imaging data. Due to the size of some plaque components being only a couple of millimeters, we argue that training a neural network on smaller image patches will perform better than a classifier based on the whole image. Besides the size of components, this decision is motivated by the observation that plaque components are not uniformly distributed throughout the whole carotid wall and that a locality-sensitive segmentation is likely to obtain better segmentation accuracy. Our model achieved good results in the segmentation of fibrous tissue but had difficulties in the segmentation of lipid and calcified tissue due to the quality of ultrasound images.

Cardiovascular disease (CVD) is one of the leading causes of death in urban areas. Carotid artery segmentation is the initial step in the automated diagnosis of carotid artery disease. The segmentation of carotid wall and lumen region boundaries are used as an essential part in assessing plaque morphology. In this paper, two types of Convolutional Neural Network (CNN) architectures are used for segmentation: U-Net and SegNet. The models used in this paper are applied on 257 ultrasound images containing a transverse section of the vessel acquired by ultrasound. Ultrasound imaging is noninvasive, completely unharming for the patient and a low-cost imaging method, but the main challenge when working with this kind of images is a very low signal to noise ratio and the process of imaging is highly dependent on the device operator. Different models are tested for various ranges of hyperparameter values and compared using different metrics. The model presented in this paper achieved over 94% Dice Coefficient for wall and lumen segmentation when trained during 100 epochs.

No abstract