Reconstruction Of Magnetic Resonance Imaging By Three-Dimensional Dual-Dictionary Learning

Magnetic resonance imaging (MRI) has been widely applied to human diseasedetection owing to its high resolution, lack of ionizing radiation and non-invasion.However, the data acquisition speed is a critical factor in multi-slice MRI applications suchas cardiac imaging or functional MRI.One solution is undersampling k-space data, which accelerates the acquisition speedbut also compromises the spatial resolution and introduces image artifacts. Thereconstruction of magnetic resonance imaging (MRI) based on compressed sensing hasbeen a hot topic in recent years because it enables accurate reconstruction fromundersampled k-space data. Our goal is to reconstruct multi-slice MRI fromundersampled k-space data employing compressed sensing (CS) theory.A novel method is proposed for multi-slice MRI reconstruction based oncompressed-sensing theory using dictionary learning. There are two aspects to improve thereconstruction quality. One is that spatial correlation among slices is implicitly employedby extending the atoms in dictionary learning from patches to blocks. In this way, spatialcorrelation among slices is fully exploited implicitly with no artificial interference. Theother is that the dictionary learning scheme is used at two resolution levels; i.e., alow-resolution dictionary is used for sparse coding and a high-resolution dictionary is usedfor image updating. Inherent correspondence between the atoms in the two dictionariesserves as effective a priori information in the updating step.The proposed algorithm is simply composed of two steps: adaptive sparse coding withdictionary of low resolution in one step, then restoring and filling in the data ofthree-dimensional k-space with dictionary of high resolution in another step.Numerical experiments were carried out on three-dimensional in vivo MR images ofbrains and abdomens with a variety of undersampling schemes and ratios. The results showthat, the proposed method achieves better reconstruction quality than conventional reconstruction methods, with the peak signal-to-noise ratio being7dB higher. Theadvantages of the dual dictionaries are obvious compared with the single dictionaryespecially for richly detailed images. Additionally, the effect of parameter variations wasanalyzed. Parameter variations ranging from50%to200%only bias the image qualitywithin3%in terms of the peak signal-to-noise ratio.