Denoising Human Heart Magnetic Resonance Diffusion Tensor Images Using Sparse Representation

Since magnetic resonance diffusion tensor imaging (DT-MRI) can provide thefunctional information of biological organization at the molecular level, it is theonly non-invasive way of fiber structure detection for in vivo tissue in the currentresearch. Today DT-MRI has been applied to the research on brain, spinal cord,liver and muscle for diagnosis of their micro-structures, which have achieved goodpreliminary results. However, the analysis of human heart functions and heart fiberreconstruction based on DT-MRI are still new research areas to be developed.There are important clinical significance and potential social economic benefits toapply DT-MRI to study the diffusion tensor characteristics and fiber structure of thepathological changes for various cardiovascular diseases.Because of noise interference, the intensity of diffusion weighted data is weakfor DT-MRI techniques. Hence the signal-to-noise ratio and contrast are lower thannormal MRI. If reliable information can not be obtained from the data, the accuracylevel of the calculated diffusion tensor is questionable, thus greatly affecting theinformation acquisition of myocardium fiber structure and reducing fiberreconstruction feasibility. Accordingly, this thesis aims towards the study of a heartDT-MRI denoising technique based on sparse representation, which is one of theprimary and difficult issues to be resolved for the development of heart DT-MRItechniques.For classical sparse denoising algorithm, some of the noise are represented asimage features and thus inducing artifacts. This negative effect is particularlysignificant in image regions with low regularization degree. Therefore, a totalvariation regularization is introduced into the sparse denoising model, and aparameter adaptive total variation based sparse representation denoising algorithmis proposed. The adaptive parameter can adjust the weight between denoising termand fidelity term according to the filtering level of each point, which allows forsuppressing the residual noise and artifacts and compensating the lost structureinformation. Thus, the denoising results can be improved compared with classicalsparse denoising algorithm.According to the multi-component characteristic of human heart DT-MRI, amulti-component image partition sparse denoising algorithm is developed. Theimage will be divided into different partitions based on the non-stationary degree,and a set of structure adaptive step atoms are designed for transition region between different partitions. Then a mixed dictionary composed of DCT basis, Haar basisand structure adaptive step atoms is constructed. Finally, the denoising isimplemented on each partition separately. This algorithm has good performance indenoising image features and non-stationary points by improving the matchingdegree between images and dictionaries.Considering the spatial correlation in cardiac DT-MRI sequence images, thisresearch proposes a3D sparse denoising algorithm for sequence images as well as aK-SVD based data training algorithm for the design of a3D structure adaptivedictionary, which enables the3D atoms have coherent spatial correlations with3Dimages. In this way, the sparse representation denoising is extended into3D spacetaking advantage of the structure similarity between neighbor slices. This algorithmshould be applied for DT-MRI sequence images denoising and it can preserve theunique fine structures of each slice effectively.Since cardiac DT-MRI images exhibit self-similarity, the study advances todevelop a structure adaptive sparse denoising algorithm on the basis of image self-similarity. A new similarity measurement with a threshold constraint is defined. Onthis basis, similar patches are searched from the image and grouped into3D data.However, this thesis implements structure adaptive3D data extracted from squarepatches with a structure adaptive window. Finally, principal component analysis isused to denoising the3D data in transform domain. This algorithm has goodperformance in denoising images with high structural redundancy. It can achieve atrade-off between image contrast and smoothness while preventing artifacts indenoising.Denoising human heart DT-MRI with sparse representation is examined andanalyzed in this research. As the basis for myocardium diffusion tensor analysis andDT-MRI denoising assessment, a level set segmentation method combining localfitting energy and gradient sensitivity energy is proposed. Myocardiums can besegmented successfully from ex vivo heart DT-MRI using this method whichenables the myocardium3D reconstruction. Experimental results show that thedenoising of heart DT-MRI can reduce the number of non-positive definite tensors,improve the accuracy of myocardial diffusion tensors and the regularization ofprincipal eigenvector fields, lower fractional anisotropy mean, increase coherenceindex mean, and finally improve the reliability of myocardium fiber reconstruction.