Liver Tumor Segmentation From CT Images Based On Fully Convolutional Neural Networks

The structure of the liver is complex and the blood vessels are rich.It is the largest organ in the abdomen of the human body.There are many types of liver lesions and the incidence rate is high.Liver cancer is one of the three deadly cancers,causing 745,000 deaths worldwide in 2012 alone.At present,computed tomography is the most com-monly used method for examining liver tumors.Tumor resection,intervention,and ra-diation are the most important treatments.Accurately knowing the size,number and lo-cation of the tumor before surgery can make a scientific and reasonable surgical plan,which is a necessary condition for successful operation.Therefore,patients with liver tumors must complete accurate segmentation of liver tumors before treatment.The following points reflect the difficulty of liver tumor segmentation.First,the organs and blood vessels around the liver are abundant,and the boundaries between normal tissues and lesions are blurred.Secondly,the CT images of liver tumors are very different,and the gray scale of the tumor is uneven.Finally,Patients with liver tumors vary greatly,and the location,shape,and size of the tumor are unpredictable.In medical diagnosis,manual segmentation has problems such as inconsistency and a lot of time.Many semi-automatic and automatic liver tumor segmentation methods have been pro-posed.The semi-automatic segmentation method requires manual intervention,relying on human subjectivity and experience.The automatic segmentation method is mainly accomplished by using the machine learning method,and it is necessary to manually extract a large number of features through massive calculations by experience.Auto-matic liver tumor segmentation is still a research difficulty and hotspot in the field of medical image processing.This paper has carried out related research on the above issues,the main research contents are as follows:1.An automatic liver tumor segmentation method based on multi-supervised full convolutional network is proposed.In-depth study of existing deep learning methods for image segmentation,including convolutional neural networks and improved full con-volutional neural networks,analyzes and compares the advantages and disadvantages of the two methods.Based on this,an automatic liver tumor segmentation method based on multi-supervised full convolution network is proposed.In a full convolutional net-work structure,the supervisory output layer is added to each layer of convolution after the third layer to guide multi-scale feature learning,better capturing local features and global features of the tumor image.The experimental results show that the mul-ti-supervised whole convolutional neural network performs better on liver tumor seg-mentation.2.Combined with the full convolutional neural network and 3D conditional random field,a new method of cascading full convolutional neural network and 3D conditional random field liver tumor segmentation is proposed.The cascading full convolutional neural network model was used to remove the influence of the surrounding environment of the liver.Finally,the segmentation image obtained by 3D conditional random field cascading full convolutional neural network was used to extract the edge information of liver tumor image to solve the liver tumor boundary.Unclear questions,complete the split.Compared with MSFCN,the segmentation result of this method is better than that of CNN and FCN,and it has high accuracy and robustness.