A Multi-contrast Mriimages Synthesis And Segmentation Methods Research For Hepatocellular Carcinoma Based On Generative Adversarial Network

Hepatocellular carcinoma(HCC) is a common primary malignant tumor of the liver,and it is also one of the cancers with high mortality.In the field of medical image processing represented by MRI,synthesizing a certain target contrast of the target anatomical structure from different contrast provides a variety of information and reliability for medical diagnosis.However,the time of MRI scanning and the physical condition of patients limit the efficiency of acquisition.At the same time,with the reason of the complex background of the liver tumor and the diversified boundaries and textures of the tumor area of the liver cancer,the precision of the MRI scanning instrument and the movement of organs cause artifacts and noises,and the use effect of the developer limit the quality of acquired data,left the facts with the clinical diagnosis misdiagnosis and missed diagnosis.Therefore,it has obvious medical significance to synthesize the supplementary target contrast of target patients from the existing contrast as the clinical diagnosis assistant of doctors.The emergence of deep learning technology has brought technological innovation to computer-aided medical analysis.This paper explores the root of image synthesis technology and the source of disadvantages of existing methods.Based on the emerging generative adversarial network,an innovative deep learning algorithm is proposed for the academic research of high-quality synthesis and segmentation of medical images represented by liver cancer MRI:We propose an generative adversarial network to meet the transfer subject between multi-contrast(T1-T2)of MRI,and work in an end-to-end way at the image level.We reserve low-frequency and high-frequency information of the image by using the adversarial loss assisted multi-stage optimization learning,build the loss of perception consistency by using the loss of perception added to the discriminator and the loss of style and content supplemented separately,and make up the deficiency of pixel by pixel loss in the difference of perception distribution by combining with the loss of cyclic consistency,so as to reserve the anatomy structure of source domain in a supervised way and learning perfectly with the perceptual pixel distribution of the contrast of the learning object.In order to integrate the different penalty(L1,L2)organically,we set the adaptive weight for the error sensitivity of the penalty function in our total loss function,so as to achieve the adaptive optimization of each stage in the generation of high-resolution image.In addition,we propose a multi-connection residual generator structure with multi-stage optimization to gradually refine medical image details.Compared with the existing technology,our method is more advanced.We have verified the T1 and T2 contrast conversion of MRI,which can help the multi contrast MRI to shorten the imaging time,improve the imaging quality,effectively assist doctors in diagnosis,and provide a new idea for super-resolution synthesis of medical images.Meanwhile,in order to verify the advanced nature of our synthetic algorithm and further expand the usability of the synthetic results,we designed a segmentation model based on generative adversarial network against the shortcomings of the traditional computer aided segmentation system of the liver cancer,and solved the segmentation difficulty caused by the complexity of the boundary and the background in the area of the liver cancer.In this paper,the double U-Net structure is used to provide the generation label supervision of the bottleneck network,and a generator model is constructed to identify the detailed edge anatomical structure of the segmented object.Finally,it is put into the training of adversarial loss.Through the effective integration of DICE loss and cross entropy loss of traditional segmentation,the accuracy evaluation of segmentation in different sizes of liver cancer is improved.