Thyroid Nodule Ultrasonic Imaging Segmentation Based on a Deep Learning Model and Data Augmentation

The segmentation of thyroid nodule ultrasonic image is a critical step for thyroid disease diagnosis. With the advent of medical big data, deep convolutional neural networks (DCNNs) have contributed to the analysis of medical image. However, there is still room for improving the accuracy of the result. In this paper, we employ several data pre-processing algorithms to amplify the feature of the original data as well as augment the whole dataset. Moreover, we use a deep learning model, improved DeepLab v3+ segmentation DCNN to achieve better training and prediction performance on thyroid nodule dataset. The results show that the dice similarity coefficient is measured to be 94.08% and accuracy is 97.91%, which reveals the advance nature of our system.