Carotid Artery and Thyroid Gland Artificial Intelligence Based Automatic Segmentation from Ultrasound Images - Qualitative Evaluation of the Segmentation Results using 3D Ultrasound Reconstructions: A Preliminary Study

Abstract

The aim of our study is finding an efficient artificial intelligence (AI) model for automatic segmentation of carotid artery from ultrasound images which may save time and lower the operator dependency during diagnosis process. A secondary objective is to provide a method also for lowering the operator dependency in carotid artery ultrasound imaging by qualitative mask evaluation and anatomy better understanding using 3D ultrasound reconstructions. A 3D ultrasound prototype based on a high frequency standard ultrasound machine and a pose reading sensor was used to acquire carotid artery and thyroid gland 2D ultrasound images. In-house-developed software applications were used for 2D image manual and automatic segmentation, 3D ultrasound reconstructions and for the 3D visualization of the scanned area. Using the 3D ultrasound reconstructions, for spatial orientation, an experienced operator created the masks remotely from the 2D ultrasound original images for three types of anatomic elements: the carotid artery wall, the lumen of the blood vessel and the thyroid gland. U-Net, U-Net++ and MultiResU-Net models were tested for automatic segmentation on a 2931 images data set. The best segmentation results in terms of IoU (Intersection over Union) and Dice metrics were obtained for the MultiResU-Net architecture: 0.9146 respectively 0.8578. The results from our preliminary study confirmed the future potential of automatic segmentation for the chosen AI models and the fact that 3D reconstructions based on 2D ultrasound images and pose reading sensors information helped the operator to obtain good quality remotely annotated ground truth (GT) masks of the carotid artery and thyroid gland.