Research On Quantitative Susceptibility Mapping Based On Deep Learning

Quantitative susceptibility mapping(QSM)is a novel magnetic resonance imaging(MRI)technique.QSM quantifies the distribution of magnetic susceptibility and shows great potential in assessing tissue content such as iron,myelin,and calcium in numerous brain diseases.The accuracy of QSM reconstruction was challenged by illposed field-to-susceptibility inversion problem,which is related to the impaired information near the zero frequency response of dipole kernel.In addition,the artifacts and noise in tissue field also have a great impact on the QSM reconstruction.Recently,deep learning methods demonstrated great capability in improving the accuracy and efficiency of QSM reconstruction.Around dipole inversion and deep learning network,we conducted two research works in this thesis:(1)Multi-Channel Networks with Dipole-Adaptive Multi Frequency Inputs for QSM reconstruction.The construction of neural networks in deep learning-based QSM methods did not take the intrinsic nature of dipole kernel into account.In this work,we proposed a Dipole Kernel-Adaptive Multi-channel CNN(DIAM-CNN)method for the dipole inversion problem in QSM.DIAM-CNN first separated the original tissue field into high-fidelity and low-fidelity components by thresholding the dipole kernel in frequency domain,and it then input the two components as additional channels into a multichannel 3D U-Net.QSM maps from Calculation of Susceptibility through Multiple Orientation Sampling(COSMOS)was used as training labels and evaluation reference.DIAM-CNN were compared with two conventional model-based methods(Morphology Enabled Dipole Inversion and improved sparse linear equation and least squares method)and one deep learning method(QSMnet).High-frequency error norm,peak signal-to-noise-ratio,normalized root mean squared error,and structure similarity index were reported for quantitative comparisons.Experiments on healthy volunteers demonstrated that the DIAM-CNN results had superior image quality to those of MEDI,iLSQR or QSMnet results.Experiment on data with simulated hemorrhagic lesion demonstrated that DIAM-CNN produced less shadow artifacts around the bleeding lesion than the compared methods.This study demonstrates that the incorporation of dipole-related knowledge into the network construction has potential to improve deep learning-based QSM reconstruction.(2)Convolutional Neural Network with Joint Denoising and Dipole Deconvolution for QSM reconstruction.The learning methods proposed since QSMnet for QSM reconstruction can be divided into two categories.In the first case,the acquired tissue field was used for network training.Under the first condition,the network had poor performance because of the artifacts and noise in the acquired tissue field.Alternatively,the simulated tissue field was used for network training.Due to the difference between the real tissue field and the simulated tissue field,the second network tended to have mundane performance.In this work,we proposed a Convolutional Neural Network with Joint Denoising and Dipole Deconvolution(JDCNN)method solving the tissue field problem.JD-CNN contains two sub-networks,namely denoising network and inversion network.Before training the inversion network,we let the real tissue field pass through the denoise network to remove the noise and artifacts.Experiments on healthy volunteers in QSMnet dataset demonstrated that the JD-CNN results had better image quality than that of QSMnet.Experiment on the external validation set demonstrated that JD-CNN had better performance compared to the above methods.This study demonstrates that the combination of denoising network and inversion network can improve the susceptibility reconstruction performance.