Research On The Algorithm Of CS-MRI Reconstruction Based On Dictionary Learning

Magnetic Resonance Imaging,MRI,one of the widely used diagnostic techniques,is carrying the advantages of low damage and high tissue contrast.The main reason for restricting the development of MRI clinical throughput and magnetic resonance imaging quality is MR image still has low scan speed.Long time of scanning will then not benefit to the Medical diagnosis,and have negative influences on patients' physical and mental health.The scanning speed can be improved in two ways,one is improving the quantity of hardware devices,the other is decreasing the amount of data collection.The second will be researched deeply in this dissertation.Compressive Sensing(CS)as new sampling theory,is compressing the data through signal sparsity.This theory can complete the signal precision recovery and decrease the amount of sample at the same time.Experiments show that the CS theory can be used in magnetic resonance image reconstruction and reduce the amount of sampling data during the process of MRI.It can be known from CS theory that the higher quantity of MRI can be achieved from the stronger image sparsity.The commonly used sparse transform in CS are wavelet transform,discrete cosine transform(DCT),finite different transform(FDT),etc.Dictionary learning(DL)method divides images into parts and definite the parts by sparse,then reconstruct the image through the integration of atoms.The advantages of DL are the ability of reaching internal structure,stronger adaptive and better consequent of the reconstructed image.Therefore,for achieving the better result,the DL method is determined to reconstruct the MRI after deeply researching MRI and CS theory.The main research in this dissertation can be concluded below:Firstly,the technique of image reconstruction based on compression sensing is studied.Compressed sensing is a kind of sampling theory to compress the data by using the sparse of the signal.It can shorten the sampling number of the data at the same time.The model of compression sensing magnetic resonance image reconstruction with total variational terms was analyzed and solved.The reconstructed results of the method were analyzed by experimental observation.Secondly,the study of compression sensing MR images based on adaptive dictionary learning is a self-adaptive sparse representation method,which uses different combinations of dictionary atoms to represent the images to be reconstructed.Compared with the fixed sparse transformation,it has the advantage of strong adaptability and can produce more sparse and more accurate sparse representation results to obtain better image reconstruction results.Finally,the pixel dispersion penalty term is added to the dictionary learning compression sensing magnetic resonance image reconstruction method.The pixel dispersion penalty term can constrain non-zero elements in the dictionary atom to form a "cluster" so that non-zero elements in the image block are more easily extracted.Experimental results show that the reconstructed method can significantly improve the peak signal to noise ratio(SNR)of the MR images,and the reconstructed images have more details and better image quality.