Computational Diffusion MRI

Diffusion imaging and brain connectivity analyses can reveal the underlying organizational patterns of the human brain, described as complex networks of densely interlinked regions. Here, we analyzed 1.5-Tesla whole-brain diffusion-weighted images from 64 participants—15 patients with behavioral variant frontotemporal (bvFTD) dementia, 19 with early-onset Alzheimer’s disease (EOAD), and 30 healthy elderly controls. Based on whole-brain tractography, we reconstructed structural brain connectivity networks to map connections between cortical regions. We examined how bvFTD and EOAD disrupt the weighted ‘rich club’—a network property where high-degree network nodes are more interconnected than expected by chance. bvFTD disrupts both the nodal and global organization of the network in both low- and high-degree regions of the brain. EOAD targets the global connectivity of the brain, mainly affecting the fiber density of high-degree (highly connected) regions that form the rich club network. These rich club analyses suggest distinct patterns of disruptions among different forms of dementia.

To classify each stage for a progressing disease such as Alzheimer’s disease is a key issue for the disease prevention and treatment. In this study, we derived structural brain networks from diffusion-weighted MRI using whole-brain tractography since there is growing interest in relating connectivity measures to clinical, cognitive, and genetic data. Relatively little work has used machine learning to make inferences about variations in brain networks in the progression of the Alzheimer’s disease. Here we developed a framework to utilize generalized low rank approximations of matrices (GLRAM) and modified linear discrimination analysis for unsupervised feature learning and classification of connectivity matrices. We apply the methods to brain networks derived from DWI scans of 41 people with Alzheimer’s disease, 73 people with EMCI, 38 people with LMCI, 47 elderly healthy controls and 221 young healthy controls. Our results show that this new framework can significantly improve classification accuracy when combining multiple datasets; this suggests the value of using data beyond the classification task at hand to model variations in brain connectivity.

Automated white matter hyperintensity (WMH) segmentation techniques for brain MRI often employ voxel-wise classifiers, trained on traditional features such as: multi-spectral MR image intensities, spatial location, texture, or shape. Recent studies show that diffusion tensor imaging (DTI) provides a measure for WMH, independent from the commonly used FLAIR images. Hence, we hypothesized that adding features derived from DTI to a voxel-wise classifier for WMH segmentation may have added value and improve segmentation results.A k nearest neighbour (kNN) classifier was implemented and trained on various combinations of features. Manual delineations of WMH were available for 20 subjects. Classifiers trained with diffusion features, such as fractional anisotropy and mean diffusivity, are compared to an equivalent classifier without diffusion features. Evaluation measures are sensitivity and Dice similarity coefficient (SI).Adding diffusion features to a kNN classifier significantly (Student’s t-test, p < 0. 0001) improved the quality of the segmentation. Depending on the chosen kNN parameters and features, improvements in sensitivity ranged from 2.4 to 13.5 % and in SI from 4.7 to 18.0 %.In conclusion, adding diffusion features derived from DTI to a voxel-wise classifier for WMH segmentation significantly improves the quality of the segmentation.

Rich information about brain tissue microstructure and composition is yielded by MRI-based measurement of the local diffusion tensor (DT) of water molecules in neural fibers, whose axons are running in myelinated fiber tracts. Diffusion tensor imaging (DTI) possesses high-dimensional and complex structure, so that detecting available pattern information and its analysis based on conventional linear statistics and classification methods become inefficient. Classification, segmentation, compression or visualization of the data could be facilitated through dimension reduction. The previously proposed methods mostly rely on complex low dimensional manifold embedding of the high-dimensional space, which are not able to deal with complex and high dimensional data. The purpose of this paper is to propose a new method for meaningful visualization of brain white matter using diffusion tensor data to map the six-dimensional tensor to a three dimensional space, employing Markov random walk and diffusion distance algorithms, leading to a new distance-preserving map for the DTI data with lower dimension and higher throughput information.

In this work, we present a new multi-parametric magnetic resonance imaging (MP-MRI) texture feature model for automatic detection of prostate cancer. In addition to commonly used imaging sequences in conventional MP-MRI, namely T2-weighted MRI (T2w) and diffusion-weighted imaging (DWI), our proposed MP-MRI texture feature model uses computed high-b DWI (CHB-DWI) and a new diffusion imaging sequence called correlated diffusion imaging (CDI). A set of texture features was calculated for both the conventional MP-MRI and new MP-MRI texture feature model. We evaluated the performance of the proposed MP-MRI texture feature model via leave-one-patient-out cross-validation using a Bayesian classifier trained on cancerous and healthy tissue samples obtained from real clinical MP-MRI datasets. The proposed MP-MRI texture feature model outperformed the conventional model (i.e., T2w+DWI) with regard to cancer detection accuracy.

Depression is a common neuropsychological consequence of stroke. The ability to predict patients at high risk of developing depressive disorders using non-invasive neuroimaging strategies has the potential to help guide treatment programs aimed to enhance functional and cognitive recovery. In this study we hypothesize that modeling the disconnection of key cortical and subcortical brain networks due to ischemic brain injury may be used to predict poststroke depression. The loss in structural connectivity was measured using diffusion-weighted MRI (dMRI) and white matter fiber tracking for 25 stroke patients (acquired 12 months after stroke) and 41 age-matched control participant. Two connectivity matrices were generated for each control participant, one with and one without the use of a manually delineated stroke lesion of a patient as an exclusion mask. A paired t-test using network-based statistics (NBS) was then performed on these connectivity matrices to determine the neural networks affected by the ischemic injury. This procedure was repeated for all stroke patients, in an independent fashion, to generate 25 disconnectivity matrices that were subsequently used in regression forest to provide a probabilistic prediction of depression. The probabilistic scores obtained from regression forests (in a leave-one-out manner) and the clinical depression scores for 25 stroke patients achieved a high positive Pearson’s correlation with ρ = 0. 78 (p < 0. 00001). This methodology shows promise as a predictive tool of poststroke depression that maybe useful for optimizing rehabilitation strategies.

Diffusion-weighted imaging and tractography offer a unique approach to probe the microarchitecture of brain tissue noninvasively. Whole brain tractography, however, produces an unstructured set of fiber trajectories, whereas clinical applications often demand targeted tracking of specific bundles. This work presents a novel, hybrid approach to fiber bundle segmentation, using spectral embedding and supervised learning. Training data of 20 healthy subjects is labeled with a parcellation-based method, and used to train support vector machine and random forest classifiers. Cross-validation was used to avoid overfitting. Results on testing data of five independent subjects show a clear improvement over unsupervised methods. Moreover, estimating the label probabilities allows to reduce the effect of outliers.

Since its introduction over a decade ago, diffusion tractography has come a long way from local, deterministic methods, over probabilistic approaches, towards global tractography. Yet, the development of tractography methods has been largely focused on single subject data, and very little on cross-population analysis and inter-subject variability. In this work, we extend global tractography with a prior on the local track orientation distribution (TOD), derived from 20 normal subjects. The proposed method is evaluated in five independent subjects. Results show that adding such prior regularizes the reconstructed track distribution, although registration errors can induce local artefacts. We conclude that atlas-guided global tractography can improve the fibre reconstruction and ultimately detect and quantify inter-subject differences in tractography.

In Diffusion Weighted Magnetic Resonance Imaging (DW-MRI) the modeling of the magnitude signal is complicated by the Rician distribution of the noise. It is well known that when dealing instead with the complex valued signal, the real and imaginary parts are affected by Gaussian distributed noise and their modeling can thus benefit from any estimation technique suitable for this noise distribution. We present a quantitative analysis of the difference between the modeling of the magnitude diffusion signal and the modeling in the complex domain. The noisy complex and magnitude diffusion signals are obtained for a physically realistic scenario in a region close to a restricting boundary. These signals are then fitted with the Simple Harmonic Oscillator based Reconstruction and Estimation (SHORE) bases and the reconstruction performances are quantitatively compared. The noisy magnitude signal is also fitted by taking into account the Rician distribution of the noise via the integration of a Maximum Likelihood Estimator (MLE) in the SHORE. We discuss the performance of the reconstructions as function of the Signal to Noise Ratio (SNR) and the sampling resolution of the diffusion signal. We show that fitting in the complex domain generally allows for quantitatively better signal reconstruction, also with a poor SNR, provided that the sampling resolution of the signal is adequate. This applies also when the reconstruction is compared to the one performed on the magnitude via the MLE.

The ensemble average diffusion propagator (EAP) obtained from diffusion MRI (dMRI) data captures important structural properties of the underlying tissue. As such, it is imperative to derive accurate estimate of the EAP from the acquired diffusion data. Taking inspiration from the theory of radial basis functions, we propose a method for estimating the EAP by representing the diffusion signal as a linear combination of 3D anisotropic Gaussian basis functions centered at the sample points in the q-space. This is in contrast to other methods, that always center the Gaussians at the origin in q-space. We also derive analytical expressions for the estimated diffusion orientation distribution function (ODF), the return-to-the-origin probability (RTOP) and the mean-squared-displacement (MSD). We validate our method on data obtained from a physical phantom with known crossing angle and on in-vivo human brain data. The performance is compared with the 3D-SHORE method of [4, 9] and radial basis function based method of [15].

In diffusion MRI, the reconstructed Ensemble Average Propagator (EAP) from the diffusion signal provides detailed insights on the diffusion process and the underlying tissue microstructure. Recently, the Simple Harmonic Oscillator based Reconstruction and Estimation (SHORE) basis was proposed as a promising method to reconstruct the EAP. However, the fitting of the basis is sensitive to noise. To solve this we propose to use the Laplacian of the SHORE basis as a natural regularization functional. We provide the derivation of the Laplacian functional and compare its effect on EAP reconstruction with that of separated regularization of the radial and angular parts of the SHORE basis. To find optimal regularization weighting we use generalized cross-validation and validate our method quantitatively on synthetic and qualitatively on human data from the Human Connectome Project. We show that Laplacian regularization provides more accurate estimation of the signal and EAP based microstructural measures.

Diffusion weighted imaging (DWI) is known to be prone to artifacts related to motion originating from subject movement, cardiac pulsation and breathing, but also to mechanical issues such as table vibrations. Given the necessity for rigorous quality control and motion correction, users are often left to use simple heuristics to select correction schemes, but do not fully understand the consequences of such choices on the final analysis, moreover being at risk to introduce confounding factors in population studies. This paper reports work in progress towards a comprehensive evaluation framework of HARDI motion correction to support selection of optimal methods to correct for even subtle motion. We make use of human brain HARDI data from a well controlled motion experiment to simulate various degrees of motion corruption. Choices for correction include exclusion or registration of motion corrupted directions, with different choices of interpolation. The comparative evaluation is based on studying effects of motion correction on three different metrics commonly used when using DWI data, including similarity of fiber orientation distribution functions (fODFs), global brain connectivity via Graph Diffusion Distance (GDD), and reproducibility of prominent and anatomically defined fiber tracts. Effects of various settings are systematically explored and illustrated, leading to the somewhat surprising conclusion that a best choice is the alignment and interpolation of all DWI directions, not only directions considered as corrupted.

Recently, super-resolution methods for diffusion MRI capable of retrieving high-resolution diffusion-weighted images were proposed, yielding a resolution beyond the scanner hardware limitations. These techniques rely on acquiring either one isotropic or several anisotropic low-resolution versions of each diffusion-weighted image. In the present work, a variational formulation of joint super-resolution of all diffusion-weighted images is presented which takes advantage of interrelations between similar diffusion-weighted images. These interrelations allow to use only one anisotropic low-resolution version of each diffusion-weighted image and to retrieve its missing high-frequency components from other images which have a similar q-space coordinate but a different resolution-anisotropy orientation. An acquisition scheme that entails complementary resolution-anisotropy among neighboring q-space points is introduced. High-resolution images are recovered at reduced scan time requirements compared to state-of-the-art anisotropic super-resolution methods. The introduced principles of joint super-resolution thus have the potential to further improve the performance of super-resolution methods.

We present and evaluate a bilateral filter for smoothing diffusion MRI fiber orientations with preservation of anatomical boundaries and support for multiple fibers per voxel. Two challenges in the process are the geometric structure of fiber orientations and the combinatorial problem of matching multiple fibers across voxels. To address these issues, we define distances and local estimators of weighted collections of multi-fiber models and show that these provide a basis for an efficient bilateral filtering algorithm for orientation data. We evaluate our approach with experiments testing the effect on tractography-based reconstruction of fiber bundles and response to synthetic noise in computational phantoms and clinical human brain data. We found this to significantly reduce the effects of noise and to avoid artifacts introduced by linear filtering. This approach has potential applications to diffusion MR tractography, brain connectivity mapping, and cardiac modeling.

Diffusion magnetic resonance imaging (dMRI) provides unique capabilities for non-invasive mapping of fiber tracts in the brain. It however suffers from relatively low spatial resolution, often leading to partial volume effects. In this paper, we propose to use a super-resolution approach based on dictionary learning for alleviating this problem. Unlike the majority of existing super-resolution algorithms, our proposed solution does not entail acquiring multiple scans from the same subject which renders it practical in clinical settings and applicable to legacy data. Moreover, this approach can be used in conjunction with any diffusion model. Motivated by how functional connectivity (FC) reflects the underlying structural connectivity (SC), we quantitatively validate our results by investigating the consistency between SC and FC before and after super-resolving the data. Based on this scheme, we show that our method outperforms traditional interpolation strategies and the only existing single image super-resolution method for dMRI that is not dependent on a specific diffusion model. Qualitatively, we illustrate that fiber tracts and track-density maps reconstructed from super-resolved dMRI data reveal exquisite details beyond what is achievable with the original data.