Block Matching Based Brain MR Image Processing Algorithms:Theory,Research And Application

Man-machine-environment system（MMESE）is an important research subject in aeronautical and astronautical area.The key component in MMESE is man,especially man’s health.With the development of magnetic resonance imaging（MRI）and digital image processing,MR image processing has become an important tool to guarantee pi-lots’and astronauts’safe.Based on the application background of MRI in aeronautics and astronautics,this thesis applies block matching（BM）in MR image processing and presents a series of algorithmic improvements.Block matching is derived from non-local means（NLM）algorithm proposed by Buades et al.This thesis focuses on the improvements and extensions of block matching.We first improve NLM and apply new algorithms in structural MRI and diffusion MRI de-noising.We then take the advantage of block matching in locating image correlated infor-mation,and leverage block matching to improve orientation distribution function（ODF）estimation and voxel-based morphology（VBM）.The proposed algorithmic improvements have strong application backgrounds.First,denoising is an important pre-processing al-gorithm and provides a good basis for various following algorithms.Second,successful tractography relies on accurate ODF estimation,and the brain connectome revealed by whole brain tractography provides valuable pathological evidences for the diagnosis of brain damages.Third,VBM has become a standard tool in the study of brain damages.The main contributions of this thesis is reflected in six points:1.A novel denoising algorithm,named collaborative block matching（CBM）,was proposed to remove the noise in structural MR images.CBM leverages both inner-image and inter-image similarity information to conduct effective noise removal.Compared with the classical NLM algorithm only utilizing image self-similarity information in denoising,CBM introduces additional inter-image similarity information and obtains improved de-noising performance.Moreover,for the unique structures in an image,CBM significantly increases the chance of locating matching structures by extending its search volume to a set of co-denoising images,and overcomes the limitation of“rare patch effect”in NLM.The experiments on synthetic data show that CBM algorithm improves the denoising per-formance both qualitatively and quantitatively.2.The NLM algorithm designed for noise removal in-space（i.e.,the spatial space）was extended to the additional-space（i.e.,the space of wavevectors）for improved de-noising performance.Diffusion data live in a combined space consisting of the-space and the-space.However,the denoising target for the classical NLM algorithm is mainly2D and 3D scalar images.Directly applying NLM in the noise removal of diffusion MRI is lack of consideration of the valuable information in-space.The proposed method,named-space BM（XQ-BM）,transfers the patch matching to-space and dramatical-ly extends the search volume to both-space and-space.XQ-BM utilizes azimuthal equidistant projection and rotation invariant features to conduct effective-space patch matching.Synthetic data experimental results show that XQ-BM significantly enhances the PSNR value and preserves the valuable edge information.3.A novel reorientation method was proposed to reorientate diffusion signals with-out involving diffusion model fitting.Traditional methods designed for this purpose in-volve diffusion model fitting in the reorientation procedure.The diffusion model fitting always introduces a large amount of computation and causes fitting errors.Moreover,after the reorientation,the information of diffusion signal is lost and following algorithms can only process reoriented diffusion models.The proposed reorientation method is an analytical solution based on Funk-Radon transformation,and directly performs on dif-fusion signals.Therefore,the proposed method avoids the diffusion model fitting and overcomes the limitations in traditional methods.The experimental results on synthetic data and real data demonstrate that the proposed method can not only accurately reorien-tate the diffusion signals,but also significantly increase the computational speed.4.A novel diffusion data preprocessing method was proposed to improve the ODF estimation.Promising ODF estimation replies on the diffusion data with high SNR and sufficient gradient samples.However,existing work only focuses on the noise removal,and is lack of consideration for the angular resolution enhancement.The proposed method utilizes multiple reference datasets to help improve the ODF estimation in the target dataset,and integrates both edge-preserving noise removal and angular resolution en-hancement in a unified framework.Specifically,block matching is first used to collec-t target and reference diffusion signals.The proposed model-fitting-free reorientation method then reorientates the reference diffusion signals.Finally,target and reorientat-ed reference diffusion signals are transferred to the sparse representation framework for ODF estimation.Through synthetic data experiments,the proposed method is proven to be effective in recovering coherent and clean ODFs.5.Block matching and non-parametric permutation test were introduced to improve the robustness of voxel-based morphology（VBM）.Voxel-based statistic and Gaussian assumption based-test are applied in traditional VBM algorithms.However,regis-tration errors,limited samples and the Gaussian assumption cause unreliable statistical conclusion in VBM.The proposed robust VBM algorithm utilizes block matching to re-duce registration errors and increase the sample size,simultaneously.Moreover,non-parametric permutation test is introduced to avoid the unreliable Gaussian assumption.In the algorithm pipeline,we further integrate exact-value,clustering analysis,multivariate Hotelling² statistic and resampling-based multiple comparison correction to optimize the algorithm performance.Synthetic data experimental results show that our method significantly reduces the registration errors and improves the convergence speed.6.The proposed four BM-based brain MR image processing algorithms were ap-plied in real applications.For T₁MRI denoising,CBM algorithm has been proven to be effectively in preserving the edge information.For DWI denoising,XQ-BM algorithm has shown remarkable performance in reducing the spurious fiber orientations caused by noise disturbance.For fiber orientation estimation,the proposed accurate ODF estima-tion algorithm has shown promising performance in reconstructing coherent and clean fiber orientations.Through the application study in real data,the proposed robust VBM algorithm has proven to obtain a reliable statistical conclusion when using a dataset with small sample size.The research presented in this thesis will improve the brain MRI processing algo-rithms used in aeronautics and astronautics.Our contribution has important significances in guaranteeing the safe and healthy of flight personnel.