Segmentation of Breast Lesions in Ultrasound Images Based on Swin-Unet

Medical image segmentation plays a key role in the early detection and diagnosis of breast cancer. Based on the Swin-Unet model in this paper, combined with the Multi-Scale Dilated Fusion Attention (MDFA) module and the Simple, Parameter-Free Attention Module (SimAM) module, improvements and optimizations have been made for the breast cancer image segmentation task. The BUSI dataset is employed, which offers abundant breast cancer image samples to assist in evaluating the effect of the model in practical applications. The MDFA module enhances the model's ability to capture features of various scales through multi-scale dilated convolution techniques, overcoming the deficiencies of traditional convolutions when dealing with complex lesion areas. The SimAM module improves the recognition ability of key features by simulating the activation process of neurons and calculating the similarity between feature points and their expected states, further enhancing the accuracy of the segmentation results. Experimental results indicate that, compared with mainstream models such as FCN, UNet, SegNet, and ENC-Net, the improved Swin-Unet demonstrates a significant improvement in performance on the BUSI dataset. The introduction of the MDFA and SimAM modules endows the model with higher precision and robustness when handling breast cancer images. This research not only validates the effectiveness of the improved model but also provides a new direction for the application of deep learning in future breast cancer image segmentation tasks.