Research On Automatic Segmentation Method Of Liver CT Image Based On Improved U-Net

Liver cancer is one of the main causes of cancer deaths,and accurate liver segmentation is an important prerequisite for cancer diagnosis and treatment.Traditionally,the task of segmentation has typically been performed manually by radiologists.However,this process is often tedious,time-consuming and easily influenced by the physician’s subjectivity,so there is an urgent need for accurate automatic liver segmentation methods to enable computer-aided diagnosis.In order to address this issue,this thesis will explore the impact of the network structure on liver segmentation based on U-Net,a very popular and suitable deep learning network framework for medical image segmentation,in turn,depending on the input dimension of the liver image（2D,3D and 2.5D）.The main research work is as follows:（1）For some difficult segmentation examples of liver CT images such as small liver regions,liver discontinuities and liver borders containing tumors,this thesis proposes an 2D adaptive multi-scale U-Net liver automatic segmentation method.A SAR-U-Net model is mainly proposed in this method.The model introduces the attention module: Squeeze and Excitation Block（SE）after each convolution operation in the traditional U-Net coding area,so that it adaptively extracts image features,filters irrelevant regions,and highlights relevant features for the segmentation task;secondly,the transition layer connecting the U-Net encoder and decoder and the output layer,which is replaced with Atrous Spatial Pyramidal Pooling（ASPP）,to obtain multi-scale image information by constructing convolution kernels with different perceptual fields through different void ratios;then,the traditional convolution blocks are replaced with residual blocks to alleviate the gradient explosion problem and prompt the network to extract more complex features.Then,a novel weighted cross-entropy loss function is proposed in order to solve the imbalance of liver image categories;finally,the validity of the method was verified through extensive experiments.（2）Aiming at the difficulty that 2D segmentation models cannot make full use of the spatial information on the z-axis of 3D medical images,this thesis proposes a3 D multi-scale U-Net liver automatic segmentation method based on depth supervision.A 3DSAI-U-Net model is mainly proposed in this method.The model first uses 3D convolution instead of 2D convolution to make full use of the information between slices of medical images;second,the standard convolution is likewise replaced using residual connections;then,the bridge connecting the coding area of the network to the decoder is replaced by a hollow Inception module with residual connections;then,a depth supervision mechanism to enable the network to make full use of the information of the image at the shallow level,ensuring a reasonable and highly accurate output at the top level.In addition,in order to better balance the segmented and unsegmented regions and reduce the proportion of false positives and false negatives,the Tversky loss function is used in the training and validation process of the network,and the hyperparameters in the loss function can be adjusted towards higher accuracy or robustness by constantly adjusting them.Finally,the validity of the method has been verified by numerous tests.（3）In response to the problem of the large number of 3D network parameters,which consume a large number of computational resources and leads to training difficulties,this thesis proposes a 2.5D lightweight multi-scale U-Net liver tumor automatic segmentation method based on a RIU-Net model.The method can segment both the liver precisely and the liver tumor efficiently.The method first applies the2.5D training approach to the CNN network,where the input to the network is in the form of inter-slice adjacencies,and at the same time generates a segmentation map corresponding to the central slice of these slice numbers;the proposed model combines the advantages of residual and Inception modules,and the network parameters are significantly reduced compared with other 2D and 3D networks,which speeds up the convergence of the model and saves After that,in order to make the model converge faster,a loss function combining cross entropy and Dice is used;finally,the method is proven to be both lightweight and efficient through extensive experiments.