Research On Reconstruction Method Of Quantitative Susceptibility Map Based On Deep Learning

Magnetic resonance imaging(MRI)is a kind of important tomography.It applies nuclear magnetic resonance theory to extract electromagnetic signals from the tissue and then reconstruct information about the tissue.MRI has become an indispensable imaging method in clinical application.Quantitative susceptibility mapping(QSM)is a novel magnetic resonance imaging technique that can measure the spatial distribution of susceptibility within tissues using the phase signal of MRI gradient echoes.QSM has shown great potential in the study of various neurological disorders,which may provide information for the diagnosis of major diseases such as Alzheimer’s disease and Parkinson’s disease.The reconstruction of QSM is a complex process.Aim to generate the final QSM image,a series of processing steps are required including deconvolution of original phase,background field removal and magnetic susceptibility inversion.The local magnetic susceptibility information is extracted from the total field after removing the background field,and the local field is deconvolution by magnetic susceptibility inversion.Deep learning has been used quite successfully in medical imaging due to its advantage over extracting deep feature.The paper proposed deep learning method to solve the problem of background field removal and magnetic susceptibility inversion in QSM image reconstruction.The main contents are as follows:First,sophisticated harmonic artifact reduction(SHARP)method leads to truncation artifact in image when truncation singular value decomposition algorithm is used to process the phase image.Multiple background field removal methods are investigated through theory analysis and experimental validation in the paper.On this basis,this paper presents a new approach called SHARPnet,which combines deep convolutional neural network with sophisticated harmonic artifact reduction method.Experimental results show that the SHARPnet method can eliminate the truncation artifacts produced by SHARP method and reconstruct high quality background removal image.Second,the paper discussed the problem about the selection of the truncation threshold of the K-space threshold method in the magnetic dipole inversion in detail by theory analysis and experiment validation.Finally,the paper proposed the model D-Net combined with deep convolutional neural network with dipole inversion model after investigating the theory of magnetic dipole inversion model.The experimental results show that D-Net method is superior to k-space threshold method in both visual and statistical quantitative evaluation.