Ultrasound Image Segmentation of Breast Tumors Based on Swin-transformerv2
Kai Huang, Zhang Yu Feng, He Meng, Feng Yue Baoping, Yao Han · 2022
Breast cancer is one of the most common cancers with a high mortality rate. Early detection is essential to reduce the risk of mortality and morbidity for breast cancer. The diagnosis of breast cancer relies heavily on the segmentation accuracy of the breast tumor region. Convolutional neural networks, a form of deep learning, have now been applied extensively to breast tumor segmentation, but they still cannot focus on global information and accurately locate features. In this paper, the swin-transformerv2-UNet(S2UNet) method for segmenting breast ultrasound images based on swin-transformerv2 is proposed. S2UNet adopts the structure of upsampling and downsampling as a whole. First, the breast ultrasound image is divided into patches, and these patches are input into the downsampling structure for feature extraction. Then, the skip connection structure is introduced to fuse the features extracted from swin-transfermerv2 in the downsampling structure with the features of the corresponding modules in the upsampling structure. Next, these features are labeled and spliced through the upsampling structure to obtain the final breast tumor image. The S2UNet method was evaluated on the breast ultrasound images dataset (BUSI) with five-fold cross-validation. The results demonstrate that compared to UNet and other advanced segmentation models (Attention-UNet, TransUNet, Swin-UNet), the S2UNet with fewer parameters shows better segmentation performance with dice similarity coefficient (DSC) of 73.26% and intersection over union (IoU) of 58.30%.