Research On Medical Image Synthesis Based On Generative Adversarial Learning

Medical image synthesis is an essential medical image processing technology that aims to construct cross-domain nonlinear mapping based on existing modal medical image data and synthesize missing medical image data.However,due to the complexity,high dimension,and large scale of medical image data,existing methods cannot effectively learn feature representation,leading to low accuracy,fidelity,and computational efficiency of medical image synthesis and challenging to deploy in practical clinical applications.Therefore,this thesis studies two methods of medical image synthesis based on generative adversarial learning to improve the capacity of learning complex high-dimensional features,increase the computational efficiency of image synthesis,and enhance the clinical application performance of medical image synthesis.The main innovations of this paper are as follows:(1)A recurrently consistent auto-encoder generative adversarial architecture based on hierarchical disentangled representation is proposed to learn the deep semantic information shared across medical images and the intrinsic deep semantic information of the source domain,utilizing the latent space alignment strategy to align the deep semantic structure levels within different latent spaces and multi-domain deep semantic knowledge,designing a disentangled representation learning strategy to separate the shallow texture detail features entangled in different feature dimensions of the latent space to achieve high-qualified medical image synthesis.(2)A two-stage lightweight generative adversarial architecture based on tensor CP decomposition is proposed to maintain the consistency between deep semantics and shallow textures of medical images.A low-rank compression network is constructed based on tensor decomposition rules to reduce the computational cost of medical image synthesis.A knowledge distillation method is proposed to transfer the dependency and distribution knowledge of high-level semantics and low-level textures of medical images to the compression architecture,ensuring that the network can synthesize accurate and high-fidelity cross-domain medical images with low computational cost.(3)The two methods proposed in this thesis were verified and analyzed experimentally on the cross-domain medical image dataset.The results show that proposed methods can effectively learn the complex and high-dimensional feature representation of medical image data and achieve medical image synthesis with accurate,high-fidelity,and high computational efficiency.The cycle-consistent auto-encoding generative adversarial architecture based on hierarchical disentangled representation learning shows sufficient feature learning ability of complex high-dimensional data.At the same time,the two-stage lightweight generative adversarial architecture based on tensor CP decomposition also offers high computational efficiency.In summary,two medical image synthesis methods are proposed based on generative adversarial learning,which can fully exploit the complex and high-dimensional features of medical image data,improving the performance and computational efficiency of medical image synthesis and providing vital support to doctors for making clinical decisions.