Research On CT Image Segmentation Of Liver Tumor Based On Fully Convolution Network

The incidence and mortality rate of liver cancer are among the highest in all kinds of cancers.Nearly one million people die of liver cancer every year,which is seriously threatening human health.Before developing a treatment plan for patients with liver cancer,doctors need to accurately understand the size,location,number and other information of the tumor,so as to accurately treat the tumor and improve the success rate of the operation.Therefore,achieving accurate segmentation of liver tumors is an important guarantee for successful operation of liver cancer patients.Due to the complex size and shape of liver tumors,the close density of tumor and liver normal tissue density,and the uneven density of liver lesions,the precise segmentation of liver tumors is challenging.A large number of workers at home and abroad have studied liver tumor segmentation,and proposed many segmentation methods,which are roughly divided into two categories:segmentation methods based on image algorithms and segmentation methods based on machine learning.The segmentation method based on image algorithm usually requires manual intervention,which is a non-automatic segmentation method.The segmentation efficiency is low and subjectively affected.The machine learning-based segmentation method mainly includes the segmentation method based on classical machine learning algorithm and the segmentation method based on deep learning.The former can automatically segment the tumor according to the artificially selected features,which improves the segmentation efficiency.The disadvantage is that the segmentation result is artificial.The set features and experience have a large impact and are computationally complex.The latter uses the convolutional neural network to automatically extract the characteristics of the tumor,which greatly improves the segmentation efficiency and achieves more accurate automatic segmentation.However,due to the problem of deep network degradation,the learning ability of the network needs to be improved.This paper designs a Deep Fully Convolutional Networks(DFCN)that can accurately and automatically segment CT images of liver tumors.DFCN uses Deep Residual Network(ResNet)as the basic network to deepen the number of layers of the fully convolutional network and improve the ability of the network to learn deep semantic information.Moreover,DFCN introduces the side output layer in the upsampling stage and combines multi-scale features to improve the details of the fully convolution network are lost.This paper proves that DFCN’s liver tumor segmentation effect is more accurate by comparing DFCN with region-based segmentation method and two methods based on fully convolution network segmentation.The fully convolutional network does not fully consider the relationship between pixels,and the segmentation result lacks spatial consistency.In this paper,the segmentation results of DFCN are optimized by the fully connected condition random field.The tumor probability map generated by DFCN provides one-potential potential energy.The color and position information between the pixels contained in the abdominal CT image is used as the binary potential energy.Finally,the average field approximation algorithm is used to iterate until the energy function value is the smallest,and the liver tumor segmentation result is obtained.The comparison experiments show that the accuracy of DFCN tumor segmentation results after the fully connected condition random field optimization is further improved.