Research On Medical Image Segmentation Algorithm Based On Transformer And Comprehensive Attention Mechanism

In recent years,deep learning has developed rapidly in the fields of natural language processing and computer vision.The role of deep learning in helping doctors diagnose diseases and pathological research can not be ignored,especially in the field of medical image segmentation.The emergence of Transformer is a good solution to the problem that CNN cannot effectively capture global information and RNN cannot realize parallel computing.However,Transformer will still generate large computing overhead in the process of batch processing videos and pictures.In this thsis,CNN is combined with several attention mechanisms,and Transformer is taken as a feature extractor of the whole network,the Gated Axial Transformer with Comprehensive Attention(CA-GAT)is proposed to solve the problems in the medical image segmentation.Firstly,in the encoder,this thsis chooses to use the axial attention mechanism instead of the self-attention mechanism in the traditional Transformer,which can achieve the same training effect as the self-attention mechanism,but greatly reduce the computational complexity.In addition,the gating mechanism is introduced to enable the model to learn accurate position bias even when the number of training data sets is small.Secondly,for medical image segmentation,the target region may be different in location,shape and scale.Therefore,three attention mechanisms of space,channel and scale are introduced to improve the ability of the model to extract fuzzy edge information.Finally,the local-global training strategy is introduced into the network,which establishes the remote dependence better and captures the global feature information better by inputting the whole image into the global branch for training.By sending the patches into the local branch,the network can learn more detailed features.The proposed model CA-GAT was validated on the skin melanoma segmentation dataset ISIC2018 and the panoramic X-ray dental segmentation dataset.In the melanoma segmentation experiment,compared with the existing mainstream models U-Net,CA-Net and Med T,the Dice coefficients of the proposed model are increased by 6.29%,0.46% and 1.20% respectively,and other indicators are also improved,which proves that CA-GAT has better segmentation performance.Furthermore,in the panoramic X-ray tooth segmentation experiment,Hausdorff distance is added to evaluate the sensitivity of the model to edge information.The experimental results show that compared with U-Net and CE-Net,the Hausdorff distance of CA-GAT decreased by 3.288 mm and 1.268 mm respectively,indicating that it has stronger edge information extraction ability.