A Deep Learning Based Study On Myelin Water Fraction Using Quantative Magnetic Resonance Imaging

Myelin water imaging（MWI）is a powerful and important magnetic resonance imaging（MRI）method for studying and diagnosing neurological and psychiatric diseases.In particular,myelin water fraction（MWF）is derived from MWI data for quantifying myelination.However,the conventional calculation of MWF based on multiple echo gradient echo sequences（mGRE）is computationally expensive,and the entire processing procedure is sensitive to the noise.In addition,the deriving process can be easily affected by the inhomogeneity of the static magnetic field（B₀）.Therefore,improving the accuracy and computational efficiency of MWF estimation is truly desired.The research of this thesis primarily focuses on improving the computational efficiency,stability and accuracy of the quantitative MWF based on mGRE sequence as the following:（1）Estimating myelin water fraction using deep artificial neural networkTo improve the computational efficiency of MWF based on the high-resolution mGRE sequence,an artificial neural network（ANN-MWF）with 6 hidden layers based on the amplitude component of T₂^*decay data has been built to estimate MWF in the current study.The experiment results showed that the result of ANN-MWF may well approximate to that reported in the literature,and may effectively reduce the computational time cost required for obtaining the resulting MWF images.In particular,ANN-MWF may effectively reduce the computational time of estimating MWF maps from 4.5 hours to less than one second,making MWF mapping a real time approach.（2）Improving the stability and accuracy of quantitative MWF mapping in high resolution using multichannel denoising convolutional neural networksTo improve the stability and accuracy of quantitative MWF mapping,a multichannel version of the denoising convolutional neural networks（MCDn CNN）to denoise the noisy mGRE data has been built so that errors and bias in the MWF quantitatifying process may be reduced in the current research.Instead of using signal-to-noise ratio（SNR）as in most of the peer studies,the current research creatively proposes the concept of noise level,and trains the denoising neural network at specific noise level according to the noise level of the data.The corresponding experiment results showed that denoising the data at a specific noise level using the corresponding specific noise level MCDn CNN may significantly improve the accuracy and stability of quantitative MWF mapping.（3）Generating myelin water fraction using an adversarial generative networkAn adversarial generative network of MWF（MWFGAN）has been developed to estimate MWF maps directly from T1W images,so that estimating MWF maps is no longer constrained to be based on a specific MWI sequence.The MWF maps derived from MWFGAN are superior to the those resulted from the conventional fitting model,in terms of image quality and accuracy statistically.In particular,conventional estimation methods frequently show structural defectiveness in their WMF maps due to B₀ inhomogeneity,which typically appears as a black hole in the maps.MWFGAN may effectively correct such structural defectiveness,thus has substantially improved the quality of WMF estimation.In conclusion,three neural networks of ANN-MWF,MCDn CNN and MWFGAN have been applied to derive MWF images in real time,improve the accuracy of MWF values,and derive MWF maps from T1-weighted images in current study,respectively.The results of the current research show the outperformance of neural network in MWF quantitative based on the T₂^*decay signal acquired by mGRE sequence.