Ultrasound thyroid nodules segmentation is a tough task because of the speckle noise, intensity heterogeneity and low contrast. In order to increase the accuracy of thyroid nodules segmentation, we present a novel framework named SG-Net. Different from other methods, we divide the segmentation problem of thyroid nodules into two steps to solve. First, a super-resolution reconstruction network is utilised to reconstruct a high-resolution version input and suppress the noise. In the second step, our proposed network perform the segmentation task. Our segmentation network includes a multi-scale input layer, several atrous spatial pyramid pooling(ASPP) modules, a U-shape convolutional backbone network with attention blocks and a our designed parallel atrous convolution(PAC) module. The multi-scale input layer paired with the ASPP module sample the input image in parallel with atrous convolutions with different dilation rates, which is beneficial to capture useful contextual information at multiple scales. The attention blocks can propagate information between the high-low level layers so that semantic features can be fully utilized for lesion segmentation. Last but not least, our proposed PAC block is beneficial to further improve the segmentation by extracting high-level semantic features from different receptive fields. Our experiments show that the Dice value of our method reaches 91.7%, and the mIoU value reaches 86.8%, which is better than most popular methods.