Comparison of Deep Learning Algorithms for Semantic Segmentation of Ultrasound Thyroid Nodules

Abstract

Ultrasound (US) imaging is used as a preliminary diagnostic tool for the detection, risk-stratification and classification of thyroid nodules. In order to perform the risk stratification of nodules in US images physicians first need to effectively detect the nodules. This process is affected due to the presence of inter-observer and intra-observer variability and subjectivity. Computer Aided Diagnostic tools prove to be a step in the right direction towards reducing the issue of subjectivity and observer variability. Several segmentation techniques have been proposed, from these Deep Learning techniques have yielded promising results. This work presents a comparison between four state of the art (SOTA) Deep Learning segmentation algorithms (UNet, SUMNet, ResUNet and Attention UNet). Each network was trained on the same dataset and the results are compared using performance metrics such as accuracy, dice coefficient and Intersection over Union (IoU) to determine the most effective in terms of thyroid nodule segmentation in US images. It was found that ResUNet performed the best with an accuracy, dice coefficient and IoU of 89.2%, 0.857, 0.767. The aim is to use the trained algorithm in the development of a Computer Aided Diagnostic system for the detection, riskstratification and classification of thyroid nodules using US images to reduce subjectivity and observer variability