Fully Convolutional Networks for Ultrasound Image Segmentation of Thyroid Nodules

Abstract

Ultrasound image segmentation plays an important role in judgement of benign and malignant thyroid nodules. Compared with the traditional convolutional neural network, the fully convolutional networks has better sparsity, higher precision and faster training speed. In this paper, we develop an 8-layer fully convolutional networks for ultrasound image segmentation of thyroid nodules, which is called FCN-Thyroid Nodules, or FCN-TN for short. We constructed a data set with 300 images to train FCN-TN. Each nodule is delineated by expert and served as ground truth for making comparison. The segmentation accuracy of 91% is obtained on the proposed network with 100 test images, which indicates that the fully convolutional networks has great potential in the field of ultrasound image segmentation of thyroid nodules.