| Liver cancer is a common malignant tumor and one of the main causes of cancer death in the world.Accurate segmentation of liver and liver tumors plays an important role in guiding doctors in disease diagnosis and surgical planning.Automated segmentation of liver tumors is of great value in clinical diagnosis and treatment.Deep learning technology is becoming more and more mature,and it have also been applied in the field of medical image processing,especially full convolution neural network,which has achieved good results in some medical image segmentation.In the application of abdominal CT image based on deep learning technology,there are two main methods: full convolutional neural networks based on two-dimensional(2D)convolution and full convolutional neural network based on three-dimensional(3D)convolution.2D convolutions can fully learn in-chip information,but ignores inter-layer information.3D convolutions can learn inter-layer information,but it is limited by its high computational complexity and high demand for GPU memory.To address these problems of the above two methods,this paper proposes a method combining the full convolutional neural network based on 2D convolution and the full convolutional neural network based on 3D convolution.The method consists of three full convolutional neural networks with different structures responsible for different functions.The first one is a full convolutional neural network based on 2D convolution,which adopts the long-connection structure of U-Net and the residual block in ResNet.Its function is to roughly segment the liver and extract the region of interest(liver region).The second is a 2.5D convolution-based full convolutional neural network with long U-Net connection structure.Its encoding branch uses DenseNet network.In the decoding branch,the pixel deconvolution layer is used to replace the deconvolution layer in the upper sampling process.The function of the network is to fully learn the information of the region of interest in the chip.The third one is full-convolution neural network based on 3D convolution,which is based on V-Net network,combined with atrous convolution.The improved V-Net network is proposed.Its function is to learn the interlayer information of regions of interest and to make the final prediction results of liver tumors.The three networks are trained separately.The training data of the third network is composed of the feature map extracted by the second network and the original image block.Finally,this paper trained on the 2017 Liver Tumor Segmentation Challenge(LiTS)Liver Tumor Data Set,and finally obtained the results of liver segmentation Dice 0.962 and liver tumor segmentation Dice 0.632 on the verification set.The results show that the segmentation effect of this method is much better than that of using only two-dimensional convolution neural network or three-dimensional convolution neural network. |
