ASPP-UNet: A New Semantic Segmentation Algorithm for Thyroid Nodule Ultrasonic Image

Thyroid nodule ultrasound image has serious noise, low contrast between different tissues, blurred boundary and irregular shape in malignant lesions. However, existing segmentation algorithms of thyroid nodule ultrasound image often have problems of coarse edge and inaccurate segmentation of small nodule. In order to segment thyroid nodules from ultrasound images more precisely, this paper proposes a new deep learning algorithm, ASPP- UN et, to achieve precise semantic segmentation of thyroid nodules from ultrasound images. In the final phase encoding ASPP module is introduced, using different void rate of atrous convolution for multi-scale features, the next sampling by using convolution instead of pooling, which can reduce the resolution of characteristic figure, can retain more detail characteristics, and use the binary cross entropy loss and damage Dice coefficient as a comprehensive loss, improve data imbalance. In the experiment, the algorithm of FeN, FusionNet, U-Net and ResUNet was repeated and compared. The experimental results showed that the ASPP-UNet model could segment thyroid nodules more precision, and the fraction of F1 score was increased from 81.7% of the original U-Net model to 90.2%. The precision was improved to 87.9%.