Research On U-Net Medical Image Segmentation Algorithm Based On Multi-Scale Feature Fusion And Densely Connection

Nowadays,the medical image segmentation plays an important role in the field of medical image processing and analysis.The accuracy of segmentation affects the final disease diagnosis directly.The traditional segmentation algorithm is limited by the complexity of medical image,and it is difficult to meet the high requirements of segmentation accuracy.As a new image processing method,convolutional neural network in the field of deep learning has been used in various kinds of medical image segmentation tasks.Many researches point out training deeper neural network can get better image segmentation results.But very deep network might lead to gradient vanishing or redundant parameters,which can affect the accuracy of image segmentation results.In order to solve these problems above,this thesis proposes a new depth neural network algorithm based on U-Net architecture,which can not only solve problems above but also improve the accuracy of medical image segmentation.Firstly,based on the concept of GoogLeNet Inception architecture,this thesis proposes a depth convolution layer for medical image segmentation.Through multi-scale feature fusion,the network can learn more complex features.Meanwhile,batch normalization layer and residual connection are used to accelerate the training process of network.Secondly,based on densely connection strategy,this thesis proposes a new type of densely connection module.Combined with the multi-scale feature fusion method,the densely connection module can learn different sizes of feature information and deepen the network layer number,which can avoid the phenomenon of gradient vanishing or explosion in training process.In order to solve the problem of training redundant parameters,convolution kernel replacement and dense connection optimization are both used to ensure the effectiveness of training process.Thirdly,based on the end-to-end network architecture of U-net,a new depth convolution neural network DIU-Net is proposed.In the beginning and final layers,the multi-scale fusion residual module based on residual connection is used,which aims to increase the width of the network and reduce the amount of parameter calculation.In the hidden layer of the network,the multi-scale fusion dense connection module densely connection layer is used,whose purpose is to increase the depth of the network and avoid gradient vanishing.In the analysis path of the whole network.the modified down-sampling module is used to reduce the mapping size of the feature image,and in the synthesis path,the modified up-sampling layer is used to adjust the mapping size of the feature image.Finally,the proposed DIU-Net model is tested on two medical image datasets,and the experimental results are evaluated and compared with traditional segmentation methods and other deep learning segmentation methods.In the segmentation dataset of lung CT images from Kaggle,the average DICE value of the test results is 0.9857,and the average AUC value is 0.9903;in the segmentation data set of retinal vessel images,the average DICE value of the test results is 0.9582,and the average AUC value is 0.9801.Compared with the results of other segmentation algorithms,the proposed DIUNet model is proved to be reliable and effective.The result shows that the proposed DIU net model can be applied to medical image segmentation tasks.