Research On Automatic Segmentation Method Of Liver CT/MRI Image Based On Generative Adversarial Network

Medical image segmentation is to extract target organs and lesions from medical digital images.In abdominal scan images,liver segmentation has important clinical applications,such as automatic liver volume measurement,liver 3D reconstruction,multi-echo nuclear magnetic inspection proton density fat grading Quantification and etc.It is difficult to segment the liver for the low contrast between the liver and neighboring organs and the changes in the texture of the diseased liver.Automatically and quickly segmenting the liver area from the abdominal scan image has always been a research hotspot.Researchers train a network for segmentation of abdominal slices based on deep neural networks,however,the segmentation results are blurry at the edges of the target organs and images of different modalities need to be trained separately.This subject took Computed Tomography(CT)and Magnetic Resonance(MR)images of the abdomen as the research objects,and trained a neural network for automatic CT/MR segmentation of the liver.The research in this paper is mainly divided into two parts: the improvement of the accuracy of liver CT image segmentation,and the conversion of liver MR image to CT image for segmentation.This topic first conducted research on improving the accuracy of abdominal CT liver image segmentation.We input abdominal CT images into Unet network for segmentation prediction,and then introduced Generative Adversarial Network(GAN)to improve the accuracy of Unet network prediction.At the same time,this subject also studied the influence of different distance constraint functions on the accuracy of liver tissue segmentation.Next,this subject used the obtained CT segmentation network to conduct liver MR image segmentation research.Liver MR images lack a large number of paired annotations.This paper applied image translation technology to convert MR images into CT images and then used CT image segmentation network for segmentation.First,the advantages and disadvantages of image translation network Cyclegan and DRIT network in segmentation of liver MR to CT image are studied..The results showed that Cyclegan’s translation results were more in line with this task,then,the impact of different structures of Cyclegan generator and discriminator in this task was studied.The results showed that the structure of Cyclegan＿G12＿D3(the generator uses a 12-layer residual network,and the discriminator uses a 3-layer network)could achieve the best results when used for segmentation of the liver MR＿T1 image conversion CT image.Finally,we used T2 sequence liver MR images to verify the liver MR segmentation method and compared with the segmentation results of the region growing method.Experimental results :(1)CT image segmentation: the Dice coefficient,Io U coefficient,pixel accuracy,relative volume error,and relative surface area error of the GAN-Unet network after the L2 distance divisor function reached 94.9%,91.3%,99.4%,0.026,0.079,all indicators were better than the traditional Unet network,and the edge of the segmentation result was smoother than the Unet network segmentation result.(2)Liver MR image segmentation: the Dice,Io U,and PA on the liver MR＿T2 image segmentation used in this paper reached 62.6%,52.3% and 95.1%.The segmentation result index and actual effect map were better than traditional region growth algorithm.This research improves the segmentation performance of liver CT images,therefore,the method can be applied to the preprocessing part of computer-aided diagnosis systems and some medical image processing tasks;At the same time,the liver MR segmentation method of this subject provides a semi-supervised segmentation of medical images by training better image translation and conversion networks,we can solve the problem of training segmentation networks due to lack of paired data sets;In addition,the researched MR image to CT image conversion result also provides an idea for medical image fusion.