Fast MR Image Reconstruction With Dictionary Learning Based Structural Information

Magnetic resonance imaging (MRI) is currently one of the most advanced techniquesin the medical imaging community. Due to the limitation of hardware equipment and thephysiological characteristics, the speed of scanning has become a bottleneck in lots of MRIapplications. Therefore, many research focus on accelerating the imaging speed ofMRI.Compressed sensing (CS) proposed in recent years, makes it possible to reconstructoriginal signals precisely from under-sampled data. Hence, applying CS to MRI (CS-MRI)can ensure the quality of reconstruction in the condition of reduce measurements data, andspeeding up the imaging process.Accurate MR images reconstruction lies in that MR images can be sparse representedin transform domains or by a set of basis functions (such as dictionary). In this thesis, wefocus on the MR image reconstruction based on dictionary learning.Synthesis dictionary based magnetic resonance image reconstruction. In synthesisdictionary model, the image of interests assumed to be represented by a linear combinationof few atoms from a given synthesis dictionary. In real world applications, it is very simpleto get a high-resolution reference image, the anatomical structure of which is similar to thetarget image. This thesis proposes a MR image reconstruction method utilizes the referenceimage based on the synthesis dictionary learning. We constitute a training set by split thereference image into overlapping patches. K-SVD algorithm is taken for synthesisdictionary learning. As there existing identical anatomical structural between the referenceimages and the target images, the learned synthesis dictionary would mirror the localstructural characteristics of target image. The objective function is established according tothe learned synthesis dictionary and the target image can be reconstructed by utilising thealternating iterative minimization algorithm. The experimental simulation outcomesmanifest that the method in this thesis not only takes advantage of the standard sparsity butalso utilizes the local structural characteristics of MR images, which is superior to theconventional method under low sampling rates. Unlike synthesis dictionary, in analysismodel, analysis operator or dictionary multiplies the image, leading to a sparse result. And the right choice of analysis operator is the key to accurate reconstruct magnetic resonanceimages. The research of analysis models still few. In this thesis, a training set is constitutedby dividing the reference image into overlapping patches, we use GOAL algorithm to learnthe analysis dictionary. At last, the target image is reconstructed by utilising the conjugategradient (CG) algorithm.