Abstract: Objective To develop a deep learning-based QSM reconstruction method for reducing artifacts to improve the accuracy of magnetic susceptibility results. Methods To eliminate artifacts caused by susceptibility interfaces with gigantic differences, we propose a multi-channel input convolutional neural network for artifact reduction (MAR-CNN) for solving the dipole inversion problem in QSM. In this neural network, the original tissue field was first separated into two components, which were subsequently imported as additional channels into a multi-channel 3D U-Net. MAR-CNN was compared with 3 conventional model-based methods, namely truncated k-space deconvolution (TKD), morphology enabled dipole inversion (MEDI), and improved sparse linear equation and least squares method (iLSQR), and with a 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. Results Experiments on healthy volunteers demonstrated that the results obtained using MAR-CNN had superior peak signal-to-noise ratio (43.12±1.19) and normalized root mean squared error (51.98±3.65) to those of TKD, MEDI, iLSQR and QSMnet. MAR-CNN outperformed QSMnet reconstruction on all the 4 quantitative metrics with significant differences (P<0.05). Experiment on data of simulated hemorrhagic lesion demonstrated that MAR-CNN produced less shadow artifacts around the bleeding lesion than the other 4 methods. Conclusion The proposed MAR- CNN for artifact reduction is capable of improving the accuracy of deep learning- based QSM reconstruction to effectively reduce artifacts.

Key words: quantitative susceptibility mapping; deep learning; convolutional neural networks; image reconstruction