An End to End Thyroid Nodule Segmentation Model based on Optimized U-Net Convolutional Neural Network

Abstract

For current clinical diagnosis of thyroid nodules, thyroid ultrasound is one of the most valuable imaging examinations to evaluate thyroid diseases. There are many improved ultrasound equipment whose imaging mechanism will cause large imaging noise, blurred borders, complicated background, which certainly bring great challenges to the nodule segmentation. As a consequence, there will be disadvantages of poor segmentation accuracy or high model complexity when using the ordinary image segmentation methods. This paper proposes an Optimized U-Net convolutional neural network model of thyroid nodule segmentation method whose structure is mainly based on U-Net model and combines the advantages of residual network. The segmentation method is also combined with the TTA (test time segmentation) method, that is, the output is the weighted average of all prediction results of the input image after enhancement. The network model trained on 544 thyroid nodule images not only achieves the end-to-end segmentation output, but also can achieve a dice coefficient of 89.50% in the final verification set.