Research On Liver Tumor CT Image Segmentation Method Based On Deep Learning

With the increase of life pressure,people's life rhythm is faster and faster,and their living habits are more and more irregular,which may lead to malignant diseases such as liver tumors,brain tumors and so on,while the incidence of liver tumors is the highest.Computed tomography(CT)is the most commonly used medical imaging technology,which can help doctors to segment liver tumors more accurately.At present,although manual segmentation is accurate,it requires a high professional level and experience of doctors,and the efficiency is very low.The semi-automatic segmentation method improves the efficiency compared with manual segmentation,but it still needs manual intervention and is vulnerable to the influence of personal subjective consciousness.Although there are some automatic segmentation methods,most of them are traditional machine learning methods,which require manual design experiments to extract features.Moreover,feature extraction requires researchers to have rich experience and spend a lot of time.In this paper,the method of liver tumor segmentation based on depth learning and its improvement are studied.Firstly,in view of the shortcomings of manual segmentation,semi-automatic segmentation and traditional automatic segmentation methods,in this paper,full convolution neural network(FCN)is used to segment liver tumors in CT images.VGG-16 is used as the basic network,the last full-connection layers are replaced by convolution layers,and then the full-convolution neural network FCN is constructed through up-sampling and information fusion operations.Experiments show that using full convolution neural network FCN to segment liver tumors in CT images can achieve good results.Then,we use U-net network to segment CT images of liver tumors,and get better results than FCN network.In order to prevent the slow convergence speed and the disappearance of gradient in the training of neural network,in this paper,a part of specification layer is added to the newly constructed DN-U-net network structure,which can improve the generalization energy of the network.In order to reduce the network parameters,reduce the size of the network model and improve the operation speed of the network model,in this paper,the common convolution in the newly constructed DN-U-net network structure is replaced by the Depthwise Separable Convolutions.Experiments show that the DN-U-net network partitioning effect is better and the network model is smaller after adding the standard layer and transforming the ordinary convolution into the deep separable convolution.Meanwhile,adding the newly proposed group normalization(GN)layer to the U-net network is more accurate than adding the traditional batch normalization(BN)layer.Finally,a traditional semi-automatic segmentation method,region growing method,is used to segment liver tumors in CT images.By comparing the experimental results of CT image segmentation with the traditional method and the depth learning method,it can be concluded that the automatic segmentation method based on depth learning is better than the traditional region growing method.