Liver Segmentation Of CT Images Based On Deep Learning

Liver cancer is one of the most common cancers in China.It has the characteristics of low predictability,rapid deterioration and high mortality.The most common two kinds of primaty liver cancers are hepatocellular carcinoma and intrahepatic cholangiocarcinoma.And these two kinds of liver cancers have lots of differences in the choice of surgical treatment and prognosis evaluation.Therefore,it is very important to achieve noninvasive early accurate classification diagnosis of hepatocellular carcinoma and intrahepatic cholangiocarcinoma.What’s more,accurate automatic liver segmentation is a necessary step to classify liver cancer through computer-aided diagnosis.This paper studies the automatic liver segmentation algorithm of CT images based on deep learning,improves the traditional segmentation algorithms,achieves the effect of effective combination with classification model,and realizes the classification research from the original CT image to hepatocellular carcinoma and intrahepatic cholangiocarcinoma.The main work of this paper is as follows:(1)This paper propose three kinds of algorithms from different directions as follows:1.By combining the dilated convolution with the structure of Unet network,the receptive domain of the network can be increased without reducing the size of the image.This means that this network has the ability to reduce the loss of feature details in the process of network training.2.We combine the Unet network with the region growing algorithm and introduce a rough trained Unet model to realize the automatic and accurate selection of initial growing seed points on abdominal CT images.3.We combine the UNET network with SLIC super-pixel algorithm.This network can effectively optimize the edge of the prediction result of Unet model.(2)We use the improved segmentation algorithm combined with the liver cancer classification model on the actual abdominal CT images which are provided by Jiangsu Provincial People’s hospital.The accuracy of the classification model which we use in the validation experiment is 90.22% in the cases :without liver cancer,hepatocellular carcinoma and intrahepatic cholangiocarcinoma.The accuracy of Unet network combined with void convolution is 39.76% in hepatocellular carcinoma,79.02% in intrahepatic cholangiocarcinoma,and the total accuracy is 58.25%.The accuracy rate of Unet network combined with region growing algorithm is 62.62% in hepatocellular carcinoma,85.27% in intrahepatic cholangiocarcinoma,and the total accuracy rate is 73.29%.The accuracy rate of Unet network combined with SLIC super-pixel algorithm is 80.12% in hepatocellular carcinoma,75.22% in intrahepatic cholangiocarcinoma,and the total accuracy rate is 77.81%.