Research Of Compressed Sensing MRI Algorithm Based On Adaptive Sparse Representation

With the rapid development of the medical image processing technique,the magnetic resonance?MR?image reconstruction has been widely applied in the diagnosis of diseases.In order to reconstruct high-quality MR images in practical medical applications,traditional approaches need to sample a large number of data,resulting in a lot of time-consuming for data sampling,and the reconstructed images are susceptible to motion artifacts and noise.Meanwhile,sampling a large amount of data requires data acquisition and storage devices to meet the high demands.Therefore,how to achieve fast and high-quality reconstruction of MR images with as little sampling data and short sampling time as possible is a key issue to be addressed urgently.The compressed sensing?CS?theory breaks through the limitations of traditional approaches and can reconstruct images accurately from a small amount of sampling data.At present,CS theory has been widely employed in MR image reconstruction,and a large number of compressed sensing magnetic resonance imaging?CSMRI?algorithms have been proposed.However,there are a few defects in some existing CSMRI algorithms,such as several finely-tuned parameters,insufficient utilization of image sparsity prior and pending to further improve the quality of the reconstructed MR image under low sampling rates.Since the adaptive sparse representation model can adaptively mine sparse priors of images and capture the finer structural information of images.Accordingly,this dissertation focuses on the research of effective CSMRI algorithms combining CS theory and adaptive sparse representation model,to reconstruct high-quality MR images from highly undersampled noisy measurements.Concretely,the research contents and innovative achievements are as follows:Firstly,to further improve image reconstruction quality of the existing CSMRI algorithms under low sampling rate,the double sparse model is constructed built on the synthesis sparse model and sparse transform model.The double sparse model is utilized to learn synthesis dictionary and transform dictionary,and then an adaptive double dictionary learning CSMRI algorithm is presented.In this algorithm,the double dictionary learning method is used to the CSMRI problem,and multiple inherent sparse priors of MR images are exploited effectively under the synthesis dictionary and transform dictionary.The proposed algorithm can reconstruct MR images from undersampled K-space noisy measurements accurately.The simulation experiments confirm the validity of the proposed algorithm.Secondly,to accelerate the image reconstruction speed and improve the image reconstruction quality,the double tight frame learning model of the horizontal and vertical gradient directions in MR image gradient domain is constructed.The constructed model is applied to the CSMRI problem,and then an adaptive double tight frame CSMRI algorithm is proposed on the basis of image gradient domain.According to the sparsity of MR images in the horizontal and vertical gradient directions,the proposed algorithm leverages adaptive double tight frames in terms of l₀ norm forms in gradient domain,to respectively execute sparse representations of the two gradient directions for images.Sequentially,the proposed algorithm can obtain more sparse priors of images in gradient domain and further improve the quality of MR image reconstruction.Because the adaptive tight frame possesses low computational complexity,the proposed algorithm speeds up image reconstruction.Experimental results demonstrate the accuracy and effectiveness of the proposed algorithm.Thirdly,to acquire sparse priors of different image components and further improve MR image reconstruction quality,a group sparsity-based tight frame learning?GSTF?model is established by fusing the group sparsity with the tight frame learning.The GSTF model is implemented by employing the adaptive tight frame and the arctangent penalty function.Then,GSTF and Total Variation?TV?are integrated into the image cartoon-texture decomposition model,and this dissertation proposes an image cartoon-texture decomposition model CSMRI?CD-MRI?algorithm,which integrates GSTF and TV.The CD-MRI algorithm respectively exploits TV to global sparsely represent MR image cartoon components and utilizes GSTF model to adaptive group sparsely represent MR image texture components.This algorithm can obtain the sparse priors of image gradient domain and group sparse priors of tight frame domain,and can achieve high-quality image reconstruction.The simulation experiments show that the proposed algorithm is effective in image reconstruction.Fourthly,to further improve the reconstruction quality of MR images,lessen the finely-tuned parameters and decrease the computational complexity of algorithms,a denoising engine based on the weighting of BM3D and FFDNet models is defined.The defined denoising engine is inserted into Regularization by Denoising?RED?framework to construct the BF-RED constraint.Then,a new CSMRI algorithm built on BF-RED constraint and Epigraph method is presented to recover MR image with high quality.The proposed algorithm develops BM3D denoising algorithm to obtain the non-local similarity and sparsity of MR images,and exploits FFDNet denoising network to receive deep priors of MR images,which can be conducive to boosting the image reconstruction quality.In order to reduce the computational complexity and avoid multiple adjusting parameters,the Epigraph method is explored to tackle the optimization problem of the proposed algorithm effectively.The simulation experiments illustrate that the validity and superiority of the prposed algorithm.Finally,to further enhance the image reconstruction quality for indwelling the phase error and amplitude noise case with K-space undersampled data of MR images,the logarithmic form adaptive tight frame and TV CSMRI?TFTV-MRI?algorithm is proposed for the phase error correction and amplitude noise removal.The TFTV-MRI algorithm leverages the adaptive tight frame in terms of the logarithmic form and TV to achieve a variety of sparse priors of MR images for promoting the image reconstruction quality and eliminating the motion artifacts and amplitude noise.Experiments prove that the presented algorithm is feasible and effective.