Multi-dictionary Learning Based On Information Entropy And Geometric Directions In MRI Reconstruction

Magnetic Resonance Imaging(MRI)is one of the important noninvasive examination methods in clinical medical imaging.However,the speed of MRI imaging data acquisition is limited by many physical conditions.During the imaging process,it is easy to move involuntarily due to the fear in the confined space in the imaging process.As a result,the motion artifact appears in the magnetic resonance imaging result,causing the doctor misdiagnosis of the patient.Therefore,imaging time and imaging quality need to be balanced during magnetic resonance imaging.Compressive sensing(CS),as an innovative method of signal collection and acquisition,it also realizes the synchronization of signal sampling and compression.Compressive sensing theory has applied to magnetic resonance imaging technology,which could effectively utilize the prior information to realize the reconstruction of magnetic resonance images by some k-space data.In this thesis,compressed sensing theory is used to segment and classify the images based on the application of magnetic resonance image reconstruction.Separate dictionary training is conducted for each type of image patches to further mine the local prior information of the image and to reconstruct the higher quality of MRI from the part k-space data,the main work of this are as follows:For the current image sparsity-based dictionary sparse reconstruction of all the image patches are represented by the same dictionary,the same dictionary can not effectively reflect the difference between different types of image patches,in view of this problem.A sparse reconstruction model based on non-local similarity image patch classification is proposed.The model performs the euclidean distance between the image patches variance and the gray value of the pixel as the metric standard to classify the image patches.By using the non-local similarity image with low rank between the patches,the dictionary training and reconstruction for each type of image patches alone,the proposed method can better mine the local feature information of the MRI image and improve the reconstruction of the magnetic resonance image under the same measurement data quality.We construct a suitable objective function and make use of the local entropy and geometric direction of the image patches according to the local features of different patches and the structural similarity between various image patches.The prior information is used to classify the image patches in multilevel.Orthogonal matching pursuit(OMP)is used to orthogonalize all the selected atoms to achieve faster reconstruction of each type of image patches.The experimental results show that the proposed algorithm has higher peak signal-to-noise ratio,structural similarity and visual effect under different sampling rates.