Study Of 3D Residual Deep Models For Medical Images Processing

In today’s massive data environment and high-efficiency performance requirements,deep learning algorithms have strong feature representation and modeling capabilities due to their multi-layer nonlinear structure,and have gradually replaced traditional algorithms in many tasks.In modern medicine,doctors mainly rely on medical imaging images of various modalities to make diagnosis and treatment plans.There are many types of medical images,such as computed tomography(CT),magnetic resonance imaging(MR),and positron emission computed tomography(PET).Medical images of different modalities can provide different information for doctors,and using deep learning algorithms to analyze and process these images can greatly improve the efficiency and accuracy of diagnosis.Limited by the lack of labeled data,the research of this paper focuses on the transformation of medical images of different modalities and the generation of synthetic images,and explores the effectiveness of using synthetic medical images to assist other medical tasks.The algorithms used in this paper are the image generation algorithms in deep learning,mainly generative adversarial network(GAN)and variational autoencoder(VAE).Besides,we use the idea of auxiliary classification generative adversarial network(ACGAN),segmentation networks UNet and residual network(ResNet)to improve the effectiveness and robustness of the network.In addition,in order to process three-dimensional medical images,all data preprocessing operations and model structures used in this paper are three-dimensional,which preserves data integrity to the greatest extent and improves algorithm performance.Specifically,this paper mainly carried out the following research work:(1)Aiming at the problem of attenuation correction error in hybrid imaging PET/MR,we use the conditional generative adversarial network(CGAN)combined with the segmentation network UNet and the residual network(ResNet)to design a deep generative model named 3D VCGAN.We utilize this model to transform abdominal multi-modal MR into synthetic CT,and then use the obtained synthetic CT to reconstruct the PET image,thereby realizing an indirect PET/MR hybrid imaging method.Compared with PET/CT,PET/MR has more advantages,but it always shows a larger standard uptake value(SUV)error.Experiments shows that our method can obtain the highest accuracy and reduce the SUV error of all assessed organs to 5% or less.(2)Aiming at the problem of multi-contrast MR image synthesis,we combine the idea of conditional generative adversarial network(CGAN),conditional variational autoencoder(CVAE)and auxiliary classification generative adversarial network(ACGAN)to design a conditional auto-encoding auxiliary classification generative adversarial network(CAEACGAN).This network combines the advantages of a variety of deep generative models to realize the one-to-many reverse transformation task from single CT to multi-modal MR images.The experimental results prove the effectiveness and robustness of our model.(3)Aiming at the challenging cross-modal MR-CT deformation registration problem,we propose an indirect cross-modal image registration and abnormal image complete method guided by synthetic CT.We use conditional auto-encoding auxiliary classification generative adversarial network(CAE-ACGAN)to realize the conversion of brain multi-modal MR images to synthetic CT images,and use the obtained synthetic CT as an intermediate medium to convert cross-modal MR-CT registration to the same-modal sCT-CT registration.In this process,we can not only get the registered MR images,but also automatically repair abnormal MR images that may be contaminated,with measured CT as the standard.