Medical Image Segmentation Method Based On Improved U-Net

Medical image segmentation methods can assist doctors to quantify and diagnose lesions faster,and give more accurate treatment plans.However,in practical clinical applications,the robustness and generalization ability of medical image segmentation models are challenging due to the differences between different disease types,different image types and different cases.Although U Network(U-Net)is a universal segmentation framework,it still has the problems of limited feature expression and inaccurate segmentation.Therefore,it is extremely challenging to construct a general model with multi-type focus segmentation.To this end,we propose DTA-UNet,a U-Net based on Dynamic Convolution Decomposition(DCD)and Triple Attention(TA).The model uses Attention U-Net as the baseline network,and realizes accurate segmentation of multiple diseases on three different image data sets.The main research contents of this paper are as follows:(1)In order to solve the problem of insufficient feature representation caused by conventional convolution in Attention U-Net,a DCD based Attention U-Net segmentation algorithm was proposed.The algorithm uses DCD to replace all conventional convolution in the encode-decoding process,exchange the number of parameters for performance improvement.The Intersection over Union(Io U),Dice Similarity Coefficient(DSC),and Precision(PRE)in the stroke segmentation dataset increase by 0.0321,0.0328,and 0.0500,respectively,while the Hausdorff Distance(HD)decreases by 0.1137.The experimental results indicate that the introduction of DCD can effectively improve the feature expression ability of Attention U-Net.(2)In order to solve the problem of insufficient extraction of global context information caused by a single jump connection of Attention U-Net,a TA based Attention U-Net segmentation algorithm was proposed,which effectively captured the long-distance dependence relationship,highlighted the lesion area and lesion boundary,and weakened the influence of irrelevant information such as background.The combination of Attention Gate(AG)and TA means that the two feature maps inputted into the AG module are modified by TA,which can strengthen the same region of interest and preserve different feature information generated for the region of interest,in order to combine with the feature information encountered in the subsequent segmentation process for the same region.In the stroke segmentation dataset,Io U,DSC and PRE increased by 0.0247,0.0225 and 0.0611,respectively,while HD decreased by 0.1102.The experimental results showed that the combination of AG and TA helped the model pay more attention to the lesion area and was more conducive to the accurate segmentation of stroke.(3)Based on the success of the above two algorithms,the above DCD and TA were introduced into Attention U-Net at the same time,and a universal disease segmentation framework DTA-UNet was proposed.It was compared with six advanced methods in three different disease segmentation tasks,namely,stroke,COVID-19 and skin lesions.All indicators achieved good results,and DSC improved 0.0431,0.0672 and 0.0161 respectively compared with U-Net,HD decreased by 0.0928,0.3947,and 0.3305,respectively.The experimental results indicate that DTA-UNet has good robustness performance.