Research On Medical Image Segmentation Networks Based On U-Net

Medical images are an important reference for clinical diagnosis,and how to quickly and accurately segment the lesion area of medical images has attracted extensive attention.The use of deep learning for image processing has become the mainstream,and medical image segmentation has become a successful example of the application of deep learning in the field of image processing because of its unique application scenario.With its unique U-shaped structure,the U-Net has achieved good performance in the field of medical image segmentation.However,there are still some problems in this network such as the accuracy is not high enough.The follow-up researchers also put forward some improved U-Net networks.Although they have improved the segmentation accuracy to a certain extent,the problems such as parameters explosion are also worthy of attention.In order to further strengthen the performance of medical image segmentation and meet the requirements of accuracy and network parameters,this thesis focuses on the medical image segmentation based on U-Net.The main contents and innovative work are summarized as follows:1)This thesis introduces U-Net and its four improved medical image segmentation networks based on U-Net,including their structure,advantages,disadvantages and performance of V-Net、Multi Res UNet、CE-Net and UNet3+ networks.2)Aiming at the problem that some existing U-Net-based improved medical image segmentation networks only pay attention to the improvement of segmentation accuracy and ignore the network calculation amount,the multiscale even convolutional attention U-Net(MECAU-Net)for medical image segmentation is proposed.This network uses 2?2 even convolution instead of 3?3convolution for feature extraction at the encoder.And inspired by the idea of the multiscale idea,4?4 even convolution is used to directly transfer the obtained information to the backbone,so as to obtain more comprehensive image information and reduce additional computational complexity.At the same time,the symmetric padding is used to solve the shift problem in the process of extracting information from even convolution kernels.In addition,the convolution block attention module is added to combine the spatial and channel attention modules after the 2?2 even convolution module,which can extract richer information without adding additional computational complexity.Finally,the simulation experiment is carried out on three classic medical image datasets.The experimental results show that the proposed MECAU-Net has better segmentation performance than other medical image segmentation networks on the basis of low complexity.3)Aiming at the problem that U-Net and its improved networks do not pay enough attention to image details,a novel improved U-Net based on multiscale encoder-decoder(MEDU-Net+)for medical image segmentation is proposed.We introduce the Goog Le Net for achieving more information at the encoder of the proposed MEDU-Net+ network,and present the multiscale feature extraction for fusing semantic information of different scales at the encoder and decoder of our proposed network.At the same time,we also introduce the single skip connection to connect the information of each layer,so that there is no need to encode to the last layer and then return the information,ensure that the feature information extracted from each layer of the encoder is not wasted.In addition,we propose a new combined loss function to extract more edge information by combining the advantages of the generalized dice and the focal loss functions.Finally,we test the proposed MEDU-Net+ network and other classic medical image segmentation networks on three classic medical image datasets.We can see from the experimental results that our proposed MEDUNet+ network has a significant improvement compared with other medical image segmentation networks.