Research On Multi-contrast Magnetic Resonance Image Denoising And Reconstruction Based On Denoising Autoencoder Prior

Multi-contrast images are collected from the same anatomical position of the patient by setting different imaging parameters,so different contrast images are highly correlated and can provide complementary information,which play a vital role in clinical diagnosis and decision-making.However,since multiple contrast images need to be acquired separately,the biggest limitation is the long imaging time.In the field of multi-contrast magnetic resonance imaging,the traditional compressed sensing based magnetic resonance imaging algorithm(CS-MRI)mainly uses joint total variation(JTV)and group wavelet sparsity to utilize the correlations between multi-contrast images to help reconstruction and attempt to eliminate artifacts caused by undersampling.However,most traditional methods have the following limitations:1)each contrast image is reconstructed separately,and less prior knowledge is considered;2)it involves a large number of complex operations,and it is computationally demanding and not very versatile.Although the emergence of deep learning theory provides new ideas for multi-contrast magnetic resonance imaging,the existing multi-contrast magnetic resonance imaging methods based on deep learning are still scarce,and they are also sensitive to noise disturbances.When the sampling patterns and acceleration factors are changed,it is necessary to retrain online,which is complicated and tedious.Therefore,this thesis combines the advantages of traditional CS-MRI mathematical models and data-driven deep learning networks,and proposes multi-noise models based on the denoising autoencoder prior.Then we successfully apply it to multi-contrast magnetic resonance image denoising and fast and accurate online multi-contrast magnetic resonance image reconstruction.The main contributions include:(1)We propose an enhanced denoising autoencoder,and the multi-contrast magnetic resonance image denoising task based on multiple noise level models verifies the feasibility and effectiveness of the network.(2)A multi-model structure with reweighted strategy at different noise levels is proposed,which can not only focus on the coarse-grained features of training data,but also capture fine-grained features including edges and organizational structure,while balancing the influence of each other.(3)Multi-contrast MR image reconstruction based on multi-noise models uses deep convolutional neural network(CNN)as the basis and embeds traditional iterative algorithms to improve the robustness and generalization of the network.Besides,only one model is needed when the sampling patterns and acceleration factors change,which greatly saves online training time.Compared with other comparison algorithms,it also achieves better results.In summary,this paper focuses on the construction of a multi-noise model with average technique.Based on the research of multi-contrast magnetic resonance image denoising,the traditional iterative algorithm is embedded to improve the network,which greatly improves the robustness and generalization of the network and achieves better magnetic resonance image quality than other comparison algorithms.