Research On Medical Image Segmentation Algorithm Based On Global Information Interactio

Medical image segmentation is to accurately classify different tissues,organs or diseased areas and other parts with special meaning from the pixel level,which plays a heart role in medical image analysis.Among them,organ segmentation is a fundamental step of medical image segmentation,which has important research significance in precision radiation therapy,visualization teaching and clinical diagnosis.Presently,with the rapid development of modern artificial intelligence technology and the improvement of computing power of hardware equipment,many researchers combine medical image segmentation with deep learning technology across disciplines in order to achieve efficient task solving,reduce the workload of medical staff and improve the accuracy and consistency of segmentation results.However,due to the characteristics of medical image and the limitations of existing algorithms,fast and accurate medical image segmentation is still a very difficult task.Therefore,this paper proposes two segmentation algorithms around the above problems,and the main research work is as follows:Firstly,the medical image segmentation algorithm based on multi-scale spatial information was proposed to solve the problems of different sizes and blurred edges of human internal organs.This algorithm combines Transformer and convolution operation to build a U-shaped architecture,taking the extraction of details and global information into account,and effectively refining the segmentation effect of different organs and edges.Moreover,the continuous convolution layer with kernel size of 3×3 is used to realize the feature dimension conversion,which provides more abundant spatial information for the input sequence of the encoder.In addition,the Cross-Attention skip connection is used to strengthen the interdependence between multi-scale feature information,thereby further improve the segmentation performance of the algorithm.Secondly,the medical image segmentation algorithm based on local details and remote dependence relationship was proposed to solve the problems that internal organs of human body are difficult to distinguish due to high similarity and adhesion of segmentation targets.By adding convolution units to construct the residual expanding layer,the algorithm can restore the resolution of feature map and strengthen the quality of segmentation prediction map.Moreover,the locality inductive bias block can be added to Swin Transformer block to make up for the lack of intrinsic inductive bias of locality and scale invariability,thereby promote the accurate prediction of complex organs and adhesion areas.In addition,absolute,relative and conditional position encoding vectors are added in order to achieve accurate positioning and refine the segmented organs,further improving the segmentation accuracy of the algorithm.Finally,the two algorithms proposed in this paper were evaluated on the Synapse abdominal CT dataset and ACDC cardiac MRI dataset respectively,and corresponding ablation experiments were conducted for the improved blocks.The experimental results show that the parameters of DSC and HD obtained by this algorithm are better than those obtained by Vi T,U-Net,V-Net,Trans Unet,Swin-Unet and other methods,and can complete the segmentation task well.