Research On Liver CT Images-assisted Diagnosis Based On Deep Learning

Liver cancer is a disease with extremely high morbidity and mortality in the world,which seriously threatens human life and health.In the clinical diagnosis of liver cancers,the main treatment method is liver tumor resection.Computed tomography(CT)technology scans the abdomen to image the area containing the liver tumor.The doctor can analyze and outline the liver and liver tumors based on subjective experience,and obtain key information such as the location,size,and volume of the liver tumor.However,as the number of data increases,the manual delineation method is time-consuming and labor-intensive.Therefore,it is of great research significance to design an automatic segmentation algorithm for liver tumors.At present,there are many difficulties in the automatic segmentation of liver CT images.For example,the feature diversification problem caused by the large differences in the shape and size of liver tumors in different patients;the blurred segmentation boundary problem caused by the similar gray values of the liver and adjacent organs and tissues;liver tumors account for a small proportion of liver CT images,which leads to the problem of under segmentation in the scenario of multiple small tumors.Because of the above problems,this paper adopts the deep learning method to study the automatic segmentation of liver tumors.In this paper,a liver segmentation algorithm based on the improved U-Net is designed.First,replace the residual learning module is used to replace the series convolution operation in the network to improve the feature extraction ability of the network;second,a hybrid atrous convolution module is added to the U-Net encoding path to expand the receptive field without losing the resolution of the feature map,and enhance the ability to perceive liver features of different sizes;finally,for the problem of data imbalance,a hybrid loss function weighted by Tevrsky loss function and cross-entropy loss function is designed,which effectively reduces false positives.Through experiments on Liver Tumor Segmentation Challenge(Li TS)dataset,the results of Liver Segmentation by this method exceed those of other algorithms in Dice coefficient,Recall,and Precision.Considering the interference of other organs and tissues in liver tumor segmentation,a two-stage liver tumor segmentation method is designed.Based on the results of the liver segmentation network,the liver region in the CT image is intercepted,and liver tumor segmentation is performed on the image that only contains the liver region.Aiming at the problem of feature learning of small-sized tumor regions,this paper proposes an attention-based liver tumor segmentation network.First,an image pyramid is introduced on the input side of the network,which effectively alleviates the loss of feature information of small targets.Second,coordinate attention is added in the feature extraction stage to encode the channel and position information of the feature map,which helps the network to capture feature information with a stronger representation ability.Finally,a deep supervision module is added to the decoding path of the segmentation network to enhance the ability of the network to discriminate key information.Experiments are carried out on the Li TS dataset for the segmentation of liver tumors in this paper,and the segmentation results are improved to a certain extent compared with other methods.