Research On Image Synthesis Technology Of PET Based On Generative Adversarial Networks

The clinical diagnosis of Alzheimer disease is mostly uses brain magnetic resonance(MR)image and Positron Emission Computed Tomography(PET)image,but it will take more time and effort to obtain these two modal images in the meantime.If we can synthesize the PET image from MR image through the deep learning technology,it will bring much convenience to patients and clinical diagnosis.With the rapid development of deep learning,generative adversarial network has been widely used in the task of medical image synthesis.Therefore,this paper mainly studies the PET image synthesis algorithm based on generative adversarial network,and the goal is to synthesize PET image from MR image.In view of the existing medical image synthesis algorithms,which have the problems of blurring and low Peak Signal-to-Noise Ratio(PSNR),the main work is as follows:From the perspective of optimizing the network structure,a PET image synthesis algorithm based on improved pix2pix is proposed.The improved residual inception module is introduced into the generator to improve the feature extration ability of the generator by increaseing its depth.Attention mechanism is also introduced into the generator to make it pay more attention to the key features in the image.Multi-scale discriminators with different receptive fields are used to discriminate the synthesized and the real PET images,which further improves the synthesis ability of the generator.Finally,Multi-scale structural similarity loss is fused in the loss function to make the synthetic PET image closer to the real.From the perspective of introducing edge detector,a PET image synthesis algorithm based on improved Ea-GAN is proposed.The edge detector is introduced into the generative adversarial network,and the edge information of the image is integrated into the adversarial training through the edge detector.The structure of the discriminator in Ea-GAN is improved,and the discriminator pays more attention to the small information of brain tissue in the image by reducing its receptive field,so as to enhance the ability of the generator to synthesize small edges.The algorithm also integrates the edge loss between real PET and synthetic PET into the loss function to further improve the performance of the generator.The experimental results on ADNI dataset show that the PET image synthesis algorithm based on the improved pix2pix can learn the diversity of PET images well,and the MAE,PSNR and SSIM indexes of the PET images synthesized based on the improved pix2pix have been improved to a certain extent.The PET image synthesized based on the improved Ea-GAN has a complete brain tissue structure and is close to the real image visually.The quantitative indexes of the synthetic PET image are better than other algorithms.