Research On Liver Tumor Segmentation In CT Images Based On Deep Learning

The liver is an important organ of the human body,which determines a person’s health status.Life in modern society is stressful and irregular schedules can lead to liver related diseases such as liver cancer.Computed Tomography(CT)is the main method of liver disease detection in clinical practice.It helps doctors to separate the liver from the tumor area and facilitates follow-up treatment.In current clinical application,liver tumor isolation is usually manually separated from CT images by doctors with rich clinical experience,which is time-consuming and labor-intensive,with poor reproducibility and strong subjectivity.Although there are many automatic schemes proposed,many of them are based on traditional image processing methods or machine learning methods,which still require manual intervention and have poor robustness.Deep learning is applied to automatic segmentation of liver tumors in CT images.The specific research results are as follows:(1)In view of the disadvantages of the current traditional automatic segmentation methods,which are characterized by strong subjectivity,poor robustness and low accuracy,a segmentation method of liver tumors based on Fully Convolutional Networks(FCN),an improved FCN is proposed.Using VGG-19 network as the backbone network,the full connection layer is replaced by the convolution layer,and then pixel level and end-to-end output of segmentation results are realized by upsampling and feature fusion.Among them in order to improve the traditional FCN rough way of feature fusion,to reconstruct on sampling part of the network structure,the third time in pooling and fourth after the pooling of sampling path add a double convolution layer respectively,through the study of convolution layer enhancement characteristics of deep expression ability,can better details on multi-feature integration and optimization segmentation,Experiments show that the proposed improved FCN segmentation precision is significantly improved compared with traditional FCN.(2)In order to obtain more accurate segmentation boundaries,improve the detection rate of small tumors and thus improve the segmentation accuracy of tumors.An improved U-NET model,RAU-NET,was proposed for liver tumor segmentation in CT images.Rau-U-Net connects the residual path with deconvolution and activation function to the skip connection layer at each scale of UNET.The information in the skip connection is restricted to the image edge information and small target global information step by step,and the edge information and target global information are separated and transmitted to improve the accuracy of boundary segmentation.The dual attention module combined with channel attention and spatial attention mechanism is added to make the network focus on the transmission of tumor-related features and inhibit the transmission of other irrelevant information.Combined with Dice loss function and binary cross entropy phase loss function,Dice loss function was used to make the model focus on the similarity between the target and the gold standard,and binary cross entropy was used to stably return the corresponding gradient of each category to make the network model converge rapidly.All convolution kernels were replaced with deformable convolution kernels to enhance the feature extraction capability of the network for irregular targets and improve the accuracy of tumor boundary segmentationExperimental results showed that RAU-NET improved the accuracy of tumor edge segmentation and the detection degree of small tumor significantly compared with UNET and FCN.The Dice coefficient of liver tumor segmentation reached 86.71%,and compared with some other segmentation networks in recent years,the proposed method also showed the optimal effect on various evaluation indicators.