Application Of Deep Learning On Image Reconstruction Of Quantitative Susceptibility Mapping

For its non-invasiveness and multi-image contrast mechanism,magnetic resonance imaging(MRI)is an indispensible clinical imaging modality.Quantitative susceptibility mapping(QSM)is a noval contrast imaging approach that calculates tissue magnetic susceptibility distribution from MRI phase maps.QSM can be used to estimate calcification,iron content and venous oxygen saturation,which may provide information for diagnosing major illness.Extracting QSM images from raw MRI phase map involves several sophiscated reconstruction steps,including phase unwrapping,background field removal and dipole deconvolution.Among these,deconvolving susceptibility map from local field map is the key problem of QSM image reconstruction.Magic angle of magnetic dipole in k space implies that deconvolution is an illposed inverse problem.Up to the present,various algoriths including multi-orientation acquisition,single oriention Bayesian and non-Bayesian algorithms have been proposed to solve this problem.However,different algorithms come with different limitions.We tried to utilize deep learning,especially convolutional neural network(CNN),to solve the ill-posed inverse problem of dipole inversion in QSM.Firstly,an artifacts removal CNN model is proposed,which can be used jointly with thresholded k-space division to overcome ill-posedness.Experiments on simulated and in-vivo datasets proved our algorithm can produce high quality images.Secondly,A k-space merging algorithm is developed from the previous proposed model,which integrates the Bayesian regularization framework,sub-domian filtering in k-space and previous CNN model.It is demonstrated by experiments that proposed algorithm overperforms the state-of-the-art algorithms both statistically and visually,while also excels in reconstruction speed.