Evaluation of Performance Metrics of Thyroid Segmentation by Deep Learning Technique

Abstract

Thyroid cancer is one of the commonly seen endocrine system cancer. Thyroid nodules appear as solid or fluid-filled masses on the thyroid. In many cases the thyroid nodules do not show any symptoms and due to this it leads to the critical situation up to death.All nodules are not cancerous and so it is very important to discriminate benign from malignant nodules. For diagnosing thyroid nodule the preferred imaging modality isUltrasound imaging.Due to inhomogeneous structure segmenting thyroid gland is a great challenge.Most of the researchers have implemented semi-automatic and automatic techniques to segment the nodules. In this paper we suggest a model to segment the region of interest by modifyingthe basic U-Netmodel. The performance metrics such as true positive, accuracy, F1-measure and dice coefficient is calculated and compared with basic model.