Liver Tumor Segmentation Of Multi-phase CT Based On Dual-channel Cascaded U-Nets

Liver cancer is one of the most common malignant tumor disease in the world,whose incidence rate is even higher in China.Therefore,it is of great significance for doctors segmenting liver and tumors from CT images with the help of the developing computer-aided diagnosis technology.However,considering the various liver size and location information of different people,the complex connection between abdominal organs and tissues,the low contrast of medical image,the diversity and diffusion,fuzzy boundary and uneven density distribution of tumors,it is still a greate challenge to realize automatic segmentation of liver and tumors.This paper proposes a DC-CUNets network based on a multi-scale dilated convolution module and deep bottleneck structure to segment both liver and tumors,which effectively solve problems from the following three aspects: the fusion of multiphase image features,the scale problem of liver tumors and the optimization of network training process.The main research work and innovations of the thesis are as follows:(1)A liver tumor segmentation method based on dual channel cascaded U-Nets(DC-CUNets)is proposed.Firstly,based on the theory of cascaded FCNs(CFCNs),a two-level U-Net segmentation network is designed,which using the first layer U-Net to segment liver,and then input the segmented results as region of interest into the second layer U-Net to segment liver tumors.Secondly,the enhanced CT includes multiple groups of scanned images achieved in different time periods after the intravenous injection of contrast media,considering that different groups of images contain different image features,the second layer network is designed as a dual channel U-Nets to learn features from both arterial and venous CT images respectively,and the features from different channels are fused through feature cascade,so as to improve the segmentation accuracy of the whole liver tumors.(2)A multi-phase liver tumor segmentation method based on multi-scale DC-CUNets is proposed.The size of liver tumors under different cancer stages is various,and different sizes of tumors,especially the small size of tumors,can not be accurately reconstructed with the reduction of the feature image resolution after multi-layers.To solve the scale problem of liver tumors,a double-layer multi-scale dilated convolution module is designed to learn different features of large,medium and small scale tumors,and the multi-scale features is still fused in the output of the module.Finally,the convolution layer of the fourth module in the dual-channel U-Nets for tumor segmentation is replaced by multi-scale dilated convolution module to realize the multi-scale DC-CUNets network.(3)A multi-scale DC-CUNets liver and tumor segmentation method based on deep bottleneck architecture is proposed.The limited network depth of the original U-Net limits the performance of feature extraction.Besides,directly adding hidden layers will not only bring complex calculation of network parameters,but also lead to over fitting of small dataset segmentation networks such as medical image segmentation.In order to optimize the training process,two deep bottleneck architectures(DBAs)is proposed to replace the convolution layer in the U-Net contraction path,so as to improve the network depth and reduce the network training parameters at the same time.Finally,by fusing the DBAs with the U-Net for liver segmentation and the multi-scale dual channel U-Nets for liver tumor segmentation,the multi-scale DC-CUNets network based on DBAs is realized.The proposed method is analyzed under 40 sets of abdominal enhanced CT with high-quality liver and liver tumor annotation from Jiangsu People’s hospital.The comparison methods include singlelayer FCN network,CFCNs network and H-Dense UNet segmentation network.Experimental results show that the DC-CUNets network and the multi-scale dilated convolution module can effectively improve the overall segmentation accuracy of liver tumors,while the DBAs achieve great performance in improving the effectiveness and reliability of network training and accelerating the convergence speed of the network.