Classification of Thyroid Ultrasound Standard Plane Images using ResNet-18 Networks

Abstract

The thyroid ultrasound standard plane (TUSP) classification is quite essential for the ultrasound diagnosis of thyroid disease. The traditional method relies entirely on the ultrasonography doctor to do it manually, which is not only time-consuming and labor-intensive but also subjectively influenced by the doctor's experience and knowledge reserve. Therefore, a TUSP automatic classification method is desirable in the clinical diagnosis of thyroid ultrasound. In this paper, we proposed that using deep learning convolutional neural network (CNN) method to achieve the automatic classification of TUSP images, and the classification effect of CNN models with different structures is also compared. In our experiment, 4,509 TUSP images collected from the hospital's real data are randomly divided into 3,386 sheets as the training set and 1,123 sheets as the test set. The test set experimental results show that the 18-layer CNN model ResNet has a good performance for automatic classification of TUSP images, and the accuracy of TUSP images classification reaches 83.88%. This indicates that the deep learning method can classify TUSP images effectively, which lays a foundation for the automatic diagnosis of thyroid diseases.