Ultrasound Image Segmentation Algorithm of Thyroid Nodules Based on Improved U-Net Network
Jianhong Gan, Rongqiao Zhang · 2022
Thyroid nodule is a common clinical case, and there are different cancer risks according to the benign and malignant grades of the nodules. The shape, edge, and size of thyroid nodules are all important basis for judging the nature of thyroid nodules in clinical practice. High-precision ultrasound image segmentation of thyroid nodules plays a crucial role in auxiliary diagnosis and judgment of benign and malignant thyroid nodules. It has a high clinical significance for the efficiency and accuracy of diagnosis. In order to improve the accuracy of thyroid nodule ultrasound image segmentation, this paper proposes an improved U-Net model thyroid nodule ultrasound image segmentation method. First, an improved residual module combined with soft-pool pooling is used to form an encoder path, and downscale feature extraction is performed while reducing the loss of feature information. Then, the full-channel attention-assisted skip connection based on the polarized self-attention mechanism is used to help the model to better find the attention region at the multi-scale level and avoid the interference of similar regions to the nodule segmentation task. The experimental results show that the Dice similarity coefficient (DSC) of the proposed method can reach 0.8065, which is better than the contrast model used in the experiment, and can well achieve the segmentation of thyroid nodules.