Medical Image Segmentation Algorithm Based On Swin Transformer And CNN Hybrid Architecture

Medical image segmentation is an important research direction in medical image processing tasks.Automated segmentation of medical images can assist medical personnel in disease diagnosis.In recent years,deep learning-based segmentation methods,mainly based on U-Net,have been widely applied in medical image segmentation with the continuous development of technology.However,due to the large differences in target morphology and structure in medical images,and the scarcity of expert annotated datasets,obtaining accurate segmentation results is still a pressing technical challenge.This article proposes two different medical image segmentation algorithms based on U-Net,which take into account the characteristics of different medical segmentation tasks and achieve good segmentation performance.The specific research achievements of this article are as follows:(1)To address the lesion segmentation problem in medical image segmentation tasks,this article proposes an end-to-end medical image segmentation network called SCW-Net based on Swin Transformer and Conv Ne Xt.SCW-Net combines the advantages of Transformer and CNN,and can maintain the integrity of global information while not sacrificing the low-level feature extraction ability of CNN.The encoder of SCW-Net uses a parallel structure with Swin Transformer branch and Conv Ne Xt branch to extract global and local features from the image.The decoder of the model uses a feature fusion module to merge the feature information extracted from the two branches.To address the problem of easy loss of details in traditional upsampling methods,we use a hybrid upsampling method in the decoder of the network.Experimental results on the BUSI and i Challenge-PM datasets show that SCW-Net can more accurately segment lesion areas compared to other comparative networks.(2)For the retinal vessel segmentation task,we propose a vessel segmentation network called MSF-Net based on multi-scale feature fusion.Retinal images have the characteristics of numerous small vessels,and the morphology,thickness,and contrast levels of vessels in each region are inconsistent.To address these characteristics,MSFNet uses a multi-scale feature extraction and fusion strategy in the encoder and decoder,respectively,to combine multi-layer feature information,enabling the network to effectively capture rich morphological features and enhance the ability of network context information interaction.MSF-Net also uses a feature extraction module based on hybrid dilated convolution,and uses an attention module to focus on the channels and spatial regions that contribute most to the segmentation results,further improving the feature extraction ability of the network.Comparative experimental results on the DRIVE and CHASE＿DB1 datasets show that the segmentation method proposed in this article outperforms the current mainstream vessel segmentation algorithms and has excellent segmentation performance.