Research On The Segmentation Method Of Three-Channel Cascade SEU-Nets Based On Enhanced CT For Liver Tumor

In this thesis,a novel liver tumor segmentation algorithm is developed on the basis of enhanced CT.In reality,doctors make comprehensive judgments based on the characteristics of the three periods of enhanced CT when performing clinical diagnosis of liver tumors.Therefore,in the design of the scheme based on the U-Net network structure in this thesis,in order to fully combine the feature information between the three periods of the same slice of enhanced CT and the feature information between different slices,it is necessary to improve the accuracy of enhanced CT liver tumor segmentation.In-depth research has been carried out in many aspects.The specific research content is as follows:(1)This thesis first uses the SE structure to improve the traditional U-Net network structure to improve the feature extraction ability of the convolutional layer in the U-Net network.Secondly,based on the previous research of the research group,the overall network cascade method is used for tumor segmentation.,The input of the first liver segmentation network is the venous phase image of enhanced CT,the output is the mask of the liver area,and the input of the second liver tumor segmentation network is the mask of the liver area of the first level network,the output is the result of liver tumor segmentation.In order to effectively extract and fuse the feature information of CT images in the three periods,this thesis proposes SEU-Nets(Three-channel cascade U-Nets based on SE structure,SEU-Nets).The three-channel network is used to segment the liver tumor,and then the feature maps extracted from the three channels are cascaded feature fusion,and finally the input size is restored to the input size through the encoder,and the segmentation map of the liver tumor is finally obtained.Experiments show that compared with CFCNs,H-Dense UNet,and DC-CUNets,The threechannel cascaded SEU-Nets liver tumor segmentation network in this article can improve tumor segmentation accuracy by 1.82% compared with the best segmentation algorithm.(2)On the basis of the previous research point,it is not just a simple cascade method of feature maps obtained from three channels to perform feature fusion.This thesis considers that 3D convolution and C-LSTM can be used for multi-modal feature extraction.Specifically,3D convolution is used to fuse the feature maps of the same slice in three periods,and then C-LSTM is used to analyze the feature information between different slices.After extraction,the purpose of making full use of the multi-modal feature information of enhanced CT is finally achieved.A new multi-modal SEUNets liver tumor segmentation algorithm based on 3D convolution and C-LSTM is proposed.Experiments have shown that the use of multi-modal feature extraction methods can make the Dice coefficient of the multi-modal SEU-Nets liver tumor segmentation algorithm reach 93.19%,and the tumor segmentation accuracy is compared with the three-channel cascaded SEU-Nets liver tumor automatic segmentation network an increase of 3.07%.At the same time accuracy rate and recall rate have been improved.(3)Finally,a liver tumor segmentation algorithm based on ENGAN’s multi-modal SEU-Nets is proposed.This thesis uses multi-modal SEU-Nets as the generator of EBGAN,and the autoencoder structure as the discriminator.In the process of generating adversarial training,it can improve the liver tumor segmentation accuracy of multi-modal SEU-Nets as the generator,and because Using the method of pre-training the discriminator can also effectively accelerate the convergence speed of the network model.The addition of EBGAN’s multi-modal SEU-Nets increases the Dice coefficient of liver tumors by 2.07% compared to the multi-modal SEU-Nets without EBGAN structure.What’s more noteworthy is that the multi-modal SEU-Nets with EBGAN added the time to segment a CT slice from 0.81 s to 0.69 s,which is a decrease of 0.12 s.In summary,the multi-modal SEU-Nets with EBGAN proposed in this thesis can not only speed up the network training speed,but also make full use of the feature information between and within the slices in the three periods of the multi-modality,thereby significantly improving the segmentation accuracy of liver tumors.