Comparative Analysis of Patch-based and Full Image Methodology of Carotid Artery Plaque Semantic Segmentation in Ultrasound images

Abstract

Arterial stenosis,caused by plaque depositions within the coronary wall,can lead to stroke and even death. The main purpose of this study is to analyze two different techniques that adequately recognize and segment fibrous,calcified and lipid plaque component in the carotid wall. Correct identification of these plaque components plays key role in an evaluation of the risk of cardiovascular disease progression. Even though both methods use Convolutional Neural Network models in its core,the main difference is in the way training ultrasound imaging data is processed. U-net architecture was chosen as CNN model,since numerous previous research showed that U-net achieves respectable results for the segmentation problems on biomedical imaging data. Custom weighted loss function that combines categorical focal loss and dice loss was used. Results showed that full image segmentation attain better segmentation of the calcified plaque component,while patch-based segmentation (that splits ultrasound images into small pieces) attains better results for segmentation of fibrous and lipid plaque component. While both methods achieved good results in segmentation of background and fibrous classes,real problem was segmentation of lipid and calcified components. Unfortunately,it was showed that splitting images into patches that are too small,results in a loss of deeper features of the image.