Improved U-Net Framework for Accurate Lumen Segmentation in Cross-Sectional Ultrasound Images of the Carotid Artery

Cardiovascular disease (CVD) is among the leading causes and a severe threat to human life of death. The presence of atherosclerotic plaque in the common carotid artery (CCA) is prevalent and has a poor prognosis. Lumen segmentation of the carotid artery in transversal ultrasound images alongside segmenting the different layers and boundaries of the carotid artery in the longitudinal view are considered to be critical in detecting atherosclerosis. Segmentation is performed to identify the lumen region in case of transversal carotid artery images and plaque region in case of longitudinal carotid artery ultrasound images, which allows for the early prediction of carotid artery disease known as atherosclerosis. The segmentation of the plaque using ultrasonic imaging is critical for aiding in the diagnosis and classification of CCA. In the case of transversal carotid artery ultrasound images, lumen boundaries, and carotid wall are taken into consideration to understand the plaque morphology. Deep learning techniques were used to evaluate carotid artery segmentation and atherosclerotic plaque. In this study, semantic segmentation of CCA transversal images is performed using the three types of Convolutional Neural Network (CNN) architectures, U-Net, SegNet and improved U-Net. To evaluate the proposed model, the architecture model is trained on 2165 transversal carotid artery ultrasound images. The presented model has undergone assessment utilizing the ADAM optimizer, achieving a Dice coefficient of 87.03% and a Jaccard Index of 79.92%. Various existing models performed using different metrics for a wide range of hyperparameter values are compared with the proposed model. This system aids the development and evaluation of carotid artery segmentation and atherosclerotic plaque by segmenting the lumen region in the first place by implementing deep learning techniques.