Research On Medical Image Segmentation Based On U-Net Network

With the development of deep learning technology,excellent network models appear frequently in image processing.It has speeded up the pace of smart medicine and provided great help for doctors in the whole stage of diagnosis and treatment.However,it is difficult to obtain enough high standard data to train the model,which is a very important problem for medical image segmentation technology,and there are no common points in shape,size and texture between lesions in medical images.To solve this problem,the segmentation network should have the ability to extract detailed features.The current method of medical image segmentation is the combination of encoder and decoder.However,most types of networks are relatively shallow and have limited feature expression capabilities,,and cannot segment accurate masks in complex medical images.Through the above problems,this thesis uses the structure of encoder and decoder to further study the medical image segmentation.The following is the research content:(1)Aiming at the problem that the network reduces parameters and can extract detailed features layer by layer.According to the excellent performance of codec structure in capturing multi-scale features,this paper proposes a multi-scale segmentation model algorithm with less parameters.An efficient network architecture is improved combining with the advantages of ResidualNetwork and DenseNetwork.Compared with ResNet,the additional jump connections are added,but the model parameters used are only about half of the DenseNet parameters,which ensures that the detailed features can still be extracted layer by layer in the case of a small number of parameters.The optimized method is evaluated on the public skin lesion dataset and the local brain MRI dataset.The experiments accuracy of 0.8999 show that the parameters of this network are less than those of other networks,and the accuracy of segmentation is guaranteed.(2)Aiming at the problem that the network model is not easy to extract masks from complex scale medical images,this paper proposes a U-Net variant to improve the performance of U-Net on various segmentation tasks,which is superimposed by two U-Net network model.Using a VGG-13 pre-trained from ImageNet as the first U-Net encoder.VGG is lightweight and similar to the U-Net architecture.It is easy to connect with U-Net and allows better deep networks output.In order to capture semantic information more effectively,we add another UNet at the bottom.Connecting all decoder parts of VGG and cascaded U-Net through SE to ensure lossless detail and global feature information,while providing larger receptive field and multi-scale information.Several experimental results of 0.8459 show that the model is effective and has generalization ability.