Reconstruction Method For Intravoxel Incoherent Motion Imaging Based On Deep Neural Network

Diffusion-weighted imaging(DWI)is based on the difference in the diffusion capacity of water molecules in various tissues of the organism.By adding diffusion sensitive gradients to conventional MRI pulse sequence,the proton spin magnetic moments in water molecules that diffuse in the tissue are imposed extra dephased,resulting in additional diffusion attenuation of the magnetic resonance signal.The difference in the diffusion capacity of water molecules makes the signal attenuation in different tissues different,so that the acquired magnetic resonance signals are diffusion-weighted.Intravoxel incoherent motion(IVIM)imaging further separates the influences of diffusion and perfusion on magnetic resonance diffusion-weighted signal,and provides tissue perfusion information that previously required contrast agent injection.However,current reconstruction methods for IVIM imaging all realize by point-by-point fitting,which takes long time.Furthermore,because each pixel is independent of each other during the reconstruction,outliers are likely to appear in the reconstruction results and affect subsequent clinical diagnosis.As one of the important branches of machine learning,deep learning has shown strong potential in speech recognition,image detection,medical diagnosis and natural language processing.In this thesis,the most widely used deep convolutional neural network in deep learning is utilized for the reconstruction of IVIM imaging.The main content is as follows:1.The basic principles of magnetic resonance diffusion-weighted imaging and the background,purpose,development and current status of IVIM imaging are briefly introduced.Several traditional reconstruction methods for IVIM are described in detail,and the advantages and disadvantages of various methods are analyzed.2.A reconstruction method for IVIM imaging based on deep neural network is proposed.Simulated samples are used to train the U-Net convolutional neural network.With the help of the receptive field in the neural network,the data points in a small range are connected to each other,which greatly improves the reconstruction efficiency and avoids outliers in the reconstruction results,making the reconstruction results more conducive to clinical diagnosis.Simulated data and real human MRI data are used to test the proposed reconstruction method.Results verify the robustness and repeatability of the method.3.Influences of the simulation model parameters and the deep neural network model on the reconstruction results of IVIM imaging are discussed,including the texture intensity and signal-to-noise ratio of the simulated samples,the structure and depth of the deep neural network,and the size of the training set.Finally,based on the optimized network model and parameters,the new method is applied to the detection of human brain gliomas and edema regions.The results are comparable with T1 enhanced images.