A Review of Diffusion Tensor Magnetic Resonance Imaging Computational Methods and Software Tools

Abstract

In this work we provide an up-to-date short review of computational magnetic resonance imaging (MRI) and software tools that are widely used to process and analyze diffusion-weighted MRI data. A review of different methods used to acquire, model and analyze diffusion-weighted imaging data (DWI) is first provided with focus on diffusion tensor imaging (DTI). The major preprocessing, processing and post-processing procedures applied to DTI data are discussed. A list of freely available software packages to analyze diffusion MRI data is also provided.

1. Introduction

Whereas conventional magnetic resonance imaging (cMRI) provides methods to map the anatomy or tissue volume, diffusion-weighted imaging (DWI) of random translational water molecules offers quantitative anisotropy and orientation information that has been utilized to map the integrity or architecture of the soft tissue in the central nervous system [1-6]. Contributors to diffusion tensor anisotropy include cellular membranes, axons, myelin sheaths, and other factors [7]. Water molecular diffusion in cerebral white matter is less restricted along the axon than perpendicular to the compact bundles and hence it is termed anisotropic (see Figure 1). Gray matter is less anisotropic, while diffusion in barrier free tissue (e.g. edema, cerebrospinal fluid) is isotropic [8-10].

1.

Illustration of diffusion anisotropy using the ellipsoid representation of the single tensor model.

2. Mathematical Background

In general, DWI data are acquired on a prescribed volume (e.g. brain) by repeating the acquisition while altering the magnitude or orientation of the diffusion-sensitizing gradients. Hence, the DWI data acquired are generally multidimensional and can always be pooled as 4D data (e.g. in space x, y, z and diffusion encoding). Diffusion-weighted data are occasionally repeated in time and magnitude-averaged to enhance the signal-to-noise ratio (SNR). This data averaging can be done by the scanner software. Depending on the acquisition protocol, the DWI data may undergo model-based single or multiple diffusion tensor imaging (DTI) or model-free analysis to obtain scalar and vector metrics that can be used to map the tissue connectivity [11-13]. Currently, there are three diffusion MRI books [11-13] and several reviews on diffusion MRI [3-6, 14-19]. The sections below provide a short overview of the basics of diffusion MRI applied to whole brain human brain mapping in health and disease.

2.1 Model-based Diffusion Tensor Imaging

Mathematically, the k-th signal S_k obtained from a volume element upon applying a diffusion-weighting or b-factor b_k along the unit vector g_k can be modeled using the Gaussian mixture model (GMM) [5] as the superposition of different slowly exchanging positive-definite and symmetric tensors D_n each with a population fraction f_n:

$$S_k=S_0\sum _{n=1}^N{f}_n\text{exp}(-b_k{\widehat{g}}_k{\mathbf{\text{D}}}_{\mathbf{\text{n}}}{\widehat{g}}_k)+{\eta }_k$$ (1)

Where D_n refers to a second rank and positive definite tensor with three unique diagonal and three unique off-diagonal elements that can be represented by a 3×3 symmetric matrix [1-3, 11-13]:

$$D=\left[\begin{array}{ccc}{D}_{xx}& D_{xy}& D_{xz}\\ D_{xy}& D_{yy}& D_{yz}\\ D_{xz}& D_{yz}& D_{zz}\end{array}\right]$$ (2)

In Equation 1, S₀ refers to the reference signal obtained without diffusion-sensitization (e.g. b=0). The sum of all population fractions is unity:

$$\sum _{n=1}^N{f}_n=1$$ (3)

In Equation 1, the apparent diffusion coefficient (ADC) can be defined as:

$$AD{C}_k({\widehat{g}}_k,b_k)=-\frac{1}{b_k}\text{log}(\frac{S_k}{S_0})$$ (4)

The system of equations described in Equation 1 can be expressed as a matrix equation after defining y_k=S_k/S₀

$$y_{Kx1}=A_{K\mathit{\text{xN}}}{f}_{Nx1}+{\widehat{\eta }}_k$$ (5)

This system of equations obtained from all measurements can be solved using constrained or regularized non-linear least-squares fit methods for the unknown diffusivities for the fractions [20-27]; this analysis is the basis for single or multiple diffusion tensor imaging (DTI). For a system with a single unknown tensor (N=1), Equation 1 can be simplified by taking the logarithm to obtain a linear system of equations

$$AD{C}_k={\widehat{g}}_{k}D{\widehat{g}}_k=D_{xx}{g}_x^2+D_{yy}{g}_y^2+D_{zz}{g}_z^2+2D_{xy}{g}_x{g}_y+2D_{xz}{g}_x{g}_z+2D_{yz}{g}_y{g}_z$$ (6a)

This equation can also be written as:

$$AD{C}_k=\sum _{i=1}^{i=3}\sum _{j=1}^{j=3}{g}_k(i)g_k(j)D_{ij}$$ (6b)

A generalization of this equation assumes the presence of high order tensors (HOT) and forms the basis of generalized DTI [28-30]. To solve this linear system of equations for the 6 diffusion tensor elements, a minimum of six independent diffusion-weighted measurements are needed in addition to the reference map (S₀). In general, more than seven measurements are acquired with different diffusion b-factors and non-collinear orientations. Examples of non-collinear or uniformly distributed diffusion encoding sets are provided in Figure 2.

2.

Illustration of several uniformly distributed minimum energy and icoshderal diffusion encoding sets.

The over-determined system described by Equation 6 can be solved by least-squares and singular value decomposition (SVD) methods [28-30].

In the case where a single-b-factor is selected based on some known range of ADC, the orientation of the encoding vectors have to be uniformly distributed in three dimensional space [31-43]. The optimization of diffusion encoding schemes for white matter fibers with selected orientations such as skeletal muscle or spinal cord has been discussed by Peng and Arfanakis [44]. In the case where non-zero b-factors are acquired along with at least 15 encoding measurements for each b-factor, diffusion kurtosis imaging (DKI) methods may be used [45-46]. Additional data-driven methods such as principal component analysis (PCA) or independent component analysis (ICA) may use the moments of the measured data [47-50].

In the general case where N diffusion tensors with rank two are sought, 6*N variables need to be determined in addition to N-1 unknown population fractions subject to Equation 3 or a total of 6N+(N-1) unknowns.

Analysis of diffusion-weighted data acquired at finite SNR, angular and spatial resolutions according to the multi-tensor model may lead generally to unstable results as the exact number (N), population fractions (f), diffusion tensor orientation, and magnitude are unknown. The two-tensor case has received some attention as it appeals to determining the extent of fast and slow diffusion compartments in a voxel [51-53] or the interesting case of within or intravoxel crossing fibers [54-57]. The fast and slow diffusion tensor decoupling requires diffusion measurements with different b-factors (e.g. b=1000 and 4000 s mm^-2), while the case of multiple crossing-fibers was modeled with uniformly distributed orientations at clinically attainable b-factors (e.g b ∼ 1000 s mm^-2). In general, the two-tensor modeling problem can be solved if sufficient data are acquired at acceptable SNR using non-linear fitting approaches or can be regularized to reduce the number of unknowns by assuming cylindrical symmetry of the unknown fibers [20-27].

2.2 Model-Free Approaches

Data-driven or model-free approaches may require longer acquisition times and the diffusion-weighted data have to be acquired according to prescribed paradigms [13]. Data-driven approaches with high angular resolution diffusion (HARDI) measurements with two b-factors may use spherical harmonic decomposition (SHD) [58-63] which is based on the expansion of the measured apparent diffusion coefficient data in terms of a complete orthonormal set (e.g. spherical Legendre polynomials Y_l^m). Mathematically, this can be expressed as:

$$\mathit{\text{ADC}}(\theta ,\mathit{\varphi })={\sum }_{l=0}^{\infty }{\sum }_{m=-l}^{m=l}{a}_{lm}{\mathit{\text{Y}}}_l^m(\theta ,\mathit{\varphi })$$ (7)

It has been shown that single fibers correspond to l=0, 2 and two fibers may be represented by l=0, 2, 4, while acquisition artifacts may be represented by odd number cases (e.g. l=1, l=3). Other model-free methods such as Q-ball imaging [64-67], diffusion spectrum imaging (DSI) [68-74] and hybrid diffusion imaging (HYDI) [75] may use mathematical operations such as Funk-Radon transform (FRT) and fast Fourier transforms (FFT) to characterize the magnitude and orientation distribution function (ODF) as detailed elsewhere [13].

2.3 The Single Diffusion Tensor Model

Currently, there are several methods developed to process or model the signal obtained according to specific acquisition paradigms. Despite its simplicity, the single tensor model remains the most popular model adopted in clinical research where healthy and patient groups are compared. The single tensor model requires a minimum of seven measurements and hence it is time-efficient. Due to SNR considerations, repeated acquisition of DWI data or acquisitions reduce noise variance. Data acquisition at higher magnetic fields [76-78] increases the intrinsic signal and offers a more time-efficient approach to acquire thinner slices [79-80] that can be used to reduce partial volume averaging effects [79-80]. It has been argued in published DTI literature that at constant imaging time acquiring data with more encoding directions is better than repeated averaging of the minimally needed amount of data [43].

2.4 Diffusion Tensor Diagonalization and Scalar Metrics

The single tensor or 3×3 matrix can be diagonalized (e.g. D = EΛE^t) to obtain three orthonormal eigenvectors (E) and corresponding eigenvalues (λ₁, λ₂ and λ₃). Several scalar functions or invariants can be defined from the diffusion eigenvalues such as mean or average diffusivity (D_av), relative anisotropy (RA), fractional anisotropy (FA), linear anisotropy (CL), planar anisotropy (CP) and spherical anisotropy (CS) [14, 81]:

$$D_{av}=\frac{{\lambda }_1+{\lambda }_2+{\lambda }_3}{3}$$ (8)

$$RA=\sqrt{\frac{1}{2}}\frac{\sqrt{{({\lambda }_1-{\lambda }_2)}^2+{({\lambda }_2-{\lambda }_3)}^2+{({\lambda }_1-{\lambda }_3)}^2}}{({\lambda }_1+{\lambda }_2+{\lambda }_3)}$$ (9)

$$FA=\sqrt{\frac{1}{2}}\frac{\sqrt{{({\lambda }_1-{\lambda }_2)}^2+{({\lambda }_2-{\lambda }_3)}^2+{({\lambda }_1-{\lambda }_3)}^2}}{({\lambda }_1^2+{\lambda }_2^2+{\lambda }_3^2)}$$ (10)

$$CL=\frac{{\lambda }_1-{\lambda }_2}{{\lambda }_1+{\lambda }_2+{\lambda }_3}$$ (11)

$$CP=\frac{2({\lambda }_2-{\lambda }_3)}{{\lambda }_1+{\lambda }_2+{\lambda }_3}.$$ (12)

$$CS=\frac{3{\lambda }_3}{{\lambda }_1+{\lambda }_2+{\lambda }_3}$$ (13)

Note that RA and FA are related analytically [50]:

$$FA=\sqrt{\frac{3}{2+R{A}^{-2}}}\iff RA=\frac{FA}{\sqrt{3-2F{A}^2}}$$ (14)

In the special case of cylindrical symmetry [81] or λ₁ > λ₂ = λ₃

$$\left[\begin{array}{c}{\lambda }_1\\ {\lambda }_2\end{array}\right]=D_{av}\ast \left(1+RA\ast \left[\begin{array}{c}2\\ -1\end{array}\right]\right)$$ (15)

The principal eigenvector (e₁) is usually used in fiber tracking algorithms [12-13, 82-91] along with anisotropy measures.

2.5 Example of a DTI Acquisition Protocol

In general, DWI data are acquired covering the entire human brain at ∼ 1-3 mm axial sections with no gap and can be accomplished in 7-15 minutes. Table 1 describes a typical DTI acquisition paradigm implemented at a 3 T clinical scanner equipped with parallel imaging powerful gradient and data sampling technology. A total of 50 volumes with 8 reference (magnitude-averaged S₀) and 42 diffusion-encoded volumes at b=1000 s mm^-2 are acquired on a unit diffusion-encoding shell. Since 43 measurements are acquired, modeling the contributions from multiple tensors (e.g. N=2) should be feasible, in principle. Alternatively, the time spent to acquire the 42 encoding orientations could have been partitioned into 6 uniformly-distributed or icosahedral orientations each with 7 b-factors (0, 143, 286, 429, 571, 714, 857, 1000 s mm^-2) keeping the number of S₀ images, echo time, maximum b-factor and total scan time constant.

Table 1.

A tabulation of the acquisition parameters in a typical isotropic diffusion tensor imaging protocol.

Data Acquisition Parameter	Value
Magnetic Field	3T
Gradient Systems	4G/cm per channel
Gradient Slew Rate	200 mT/m/ms
Coil	Head/Parallel imaging
Acquistion Mode	2D-axial: Fast spin-echo-EPI
Whole Brain	Inferior-to-superior
Field-of-view	256mm × 256mm
Voxel size	2mm × 2mm × 2mm
Acquistion Matrix	128×64
Acceleration	R=2
Zero filling	k-space
Image Matrix	256×256×70slices
Diffusion b-Factors	1000 s mm-2
Reference S0	8
Signal-To-Noise Ratio	∼ 50
Encoding Directions	42 (icosahedral: alternating Icosa21b)
Echo Time (TE)	65 ms
Replication Time (TR)	12,000 ms
Scan Time	7 minutes

Acquiring DWI data using different or arbitrary diffusion encoding orientations and multiple b-factors can be rather complicated and involves trade-offs between spatial and angular resolutions. For example, acquiring data with high isotropic spatial resolution at 1mm × 1mm × 1mm and using multiple b-factors and encoding directions may be best used to model the contributions from multiple fibers in a voxel, but this may not be time-efficient and SNR will be much smaller than a whole brain 2mm × 2mm × 2mm protocol that can be acquired under 7 minutes. Acquiring data with high b-factor sensitization and high angular diffusion allows multiple tensor compartment construction (e.g. fast and slow diffusion) and modeling of crossing fibers [51-57], with a trade-off of reduced SNR and increased scan time for whole brain protocols. As illustrated in Figure 3, noisy DWI measurements lead to an overestimation of anisotropy in isotropic systems [41, 92] whereas noise reduces the estimated anisotropy in highly anisotropic systems [93]. The estimated apparent or effective tensor anisotropy in regions of crossing fibers is reduced [94].

3.

An illustration of the effect of noise and crossing fibers on the estimated diffusion anisotropy.

3.0 Overview of Computational Procedures applied to Diffusion MRI Data

Freely available MRI computational [13] and DTI software packages [95-103] summarized below require the user to upload raw data along with the data acquisition parameters that include spatial and temporal image and diffusion acquisition parameters. The preprocessing and analysis conducted by these packages depend on the software primary design and models adopted to decode the diffusion-weighted data. Figure 4 summarizes some possible steps that can be applied on the DWI data.

4.

A schematic representation of several preprocessing, processing and post-processing procedures that can be performed on the raw DWI data or processed diffusion tensor maps.

3.1 Diffusion-weighted Imaging Preprocessing

The major preprocessing steps applied on a typical DWI data are summarized in Figure 4. The acquired DWI data may undergo several steps that depend on the clinical scanner, acquisition parameters and image quality. Most clinical MRI vendors provide the data in 2D-image DICOM (digital imaging and communications in medicine) format or other special data formats which may need to be converted to the format readable by the software package of interest.

For anatomical MRI and DWI data format preparation, conversion, inspection, tissue segmentation, and visualization, Table 2 provides a list of useful web sites with freely downloadable software packages. MRIcro is a popular free and easy to use package that has a user-friendly graphical user interface (GUI) on all operating systems to convert data files from 2D (DICOM http://medical.nema.org/) to 3D (e.g. Analyze http://mayoresearch.mayo.edu/mayo/research/robb_lab/analyze.cfm) or 4D neuroimaging informatics technology initiative (NIFTI http://nifti.nimh.nih.gov/). Other packages such as AFNI, FreeSurfer, FSL, SPM offer command-line conversion capabilities that enable users to prepare DWI and batch process [104] large data sets.

Table 2.

A list of MRI image analysis tools that can be used for data conversion, preparion and tissue segmentation.

Image Processing Package	Purpose	Web Location
AFNI	Analysis of Functional NeuroImages	http://afni.nimh.nih.gov/afni/doc/misc/
BrainSuite BrainParser	Volume Segmentation And Parcellation	http://www.loni.ucla.edu/Software/BrainSuite http://www.loni.ucla.edu/Software/BrainParser
Brain Visa	Q-ball T1w-segement DTI-tracking	http://brainvisa.info/ http://www.nitrc.org/projects/brainvisa/
CRUISE	Volume Segmentation	http://www.nitrc.org/projects/toads-cruise/
JIST	Java Image Science Toolkit. Handles multi tensors	http://www.nitrc.org/projects/jist/
ICBM	Brain Atlases and software	http://www.loni.ucla.edu/ICBM/Resources/Resources_Software.shtml
ImageJ	Image Processing and Analysis in Java	http://rsb.info.nih.gov/ij/
FreeSurfer	Surface and volume parcellaltion	http://surfer.nmr.mgh.harvard.edu/
FSL BET FIRST-FAST	Brain Masking Segmentation	http://www.fmrib.ox.ac.uk/fslcourse/ http://www.fmrib.ox.ac.uk/fsl/bet2/index.html http://www.fmrib.ox.ac.uk/fsl/first/
MEDInria	Image Processing	http://www-sop.inria.fr/asclepios/software/MedINRIA/
MRIcro MRIcron Includes BET	Tools for preprocessing, visualizing and converting data types *.nii, analyze, dicom	http://www.mricro.com http://www.sph.sc.edu/comd/rorden/anatomy/index.htm
MIPAV	Medical Image Processing, Analysis, and Visualization	http://mipav.cit.nih.gov/
Slicer	Open source package for visualization and medical image computation	http://www.slicer.org/
SPM	fMRI, smoothing/VBM	http://www.fil.ion.ucl.ac.uk/spm/ http://www.fil.ion.ucl.ac.uk/spm/ext/#toolboxes

In general, the acquired DWI volume data need to be co-aligned with the reference b₀ volume and distortion correction need to be applied carefully [105,106]. Several sophistical image registration packages (e.g. AIR, ART, FNIRT) summarized in Table 3, discussed and compared elsewhere [107-108] can handle data coalignment, and spatial normalization, distortion correction [109-111]. Post registration DWI data may be subjected to simple geometric operations to reduce data storage such as cropping. In addition, brain extraction may follow to remove skull and non-brain tissues [112]. Diffusion-weighted MRI data can subsequently undergo spatial denoising (e.g. median or anisotropic edge-preserving filtering) to enhance SNR or interpolated to attain isotropic voxel dimensions as needed [113,114].

Table 3.

A list of software packages used for image registration and distortion correction.

Volume Registration Package	Purpose	Web Location
AFNI	Analysis of Functional NeuroImages	http://afni.nimh.nih.gov/afni/doc/misc/
AIR	Volume registartion Spatial Normalisation	http://bishopw.loni.ucla.edu/ http://www.loni.ucla.edu/Software/AIR
ANTs		http://www.picsl.upenn.edu/ANTS/
ART	Registration Toolbox	http://www.nitrc.org/proiects/art/
FSL FLIRT/FNIRT	Registration	http://www.fmrib.ox.ac.uk/fsl/flirt/ http://www.fmrib.ox.ac.uk/fsl/fnirt/
HAMMER		http://www.nitrc.org/projects/hammer/
IRTK		http://www.doc.ic.ac.uk/∼dr/software/
ITK		http://www.itk.org/
SPM Dartel	spatial normalization registration	http://www.fil.ion.ucl.ac.uk/spm/ http://www.fil.ion.ucl.ac.uk/spm/ext/#toolboxes

3.2 Diffusion-weighted Imaging Processing

Depending on the data acquisition, the prepared and encoded DWI data may subsequently get decoded to estimate the diffusion tensor per voxel using the acquisition encoding table (e.g. encoding orientation and b-factors). The diffusion tensor D may be diagonalized further to obtain the eigenvalues and eigenvectors. The eigenvalues may be used to obtain the mean diffusivity and anisotropy maps as defined above. Figure 5 illustrates the steps applied on some DWI data acquired using the Icosa21 encoding scheme on a single section.

5.

A pictorial illustration of the processing steps applied to a representative DWI data set using the icosa21 encoding scheme.

3.3 Diffusion Post-Processing

Most DTI quantitative packages produce diffusion tensor, FA, eigenvalue and principal eigenvector volumes from the DWI data acquired on each subject. Table 4 provides a list of free diffusion MRI software packages that have been reported in the literature. Additional MRI and diffusion MRI packages are listed on the neuroimaging tools and resources home page (http://www.nitrc.org/). Clinical MRI scanner vendor proprietary packages and commercial packages with DTI capabilities such as BrainVoyager (http://www.brainvoyager.com/index.html) and Analyze direct (http://www.analyzedirect.com/products/mridti.asp) are not discussed below.

Table 4.

A list of freely downloadable diffusion MRI software packages.

DTI Software	Purpose	Web Location
CATNAP	Coregistration, Adjustment, and Tensor-solving – a Nicely Automated Program	http://www.nitrc.org/projects/jist/ http://www.nitrc.org/plugins/mwiki/index.php/jist:MainPage#Diffusion_Tensor_Imaging
CAMINO	Diffusion MRI Toolkit.	http://web4.cs.ucl.ac.uk/research/medic/camino/pmwiki/pmwiki.php?n=Main.HomePage
DipY	Diffusion Imaging in Python	http://nipy.sourceforge.net/dipy/theory/dicom.html
DoDTI	DTI	http://neuroimage.yonsei.ac.kr/dodti/
DTIStudio DiffeoMap & ROIEditor	DTI processing, Deterministic fiber tracking	https://www.mristudio.org/wiki/installation https://www.mristudio.org/wiki/user_manual
DSIStudio	Handles QBI, DSI, DTI	http://graphics.stanford.edu/projects/dti/software/ http://dsi-studio.labsolver.org/Manual/Reconstruction
DTI-TK	tracking	http://www.nitrc.org/projects/dtitk
DTI-Query	DTI	http://graphics.Stanford.edu/projects/dti/
DTI-Toolbox	Data prep Interpolation, Smooth/Process ROI/Atlas-based segment	http://www.uth.tmc.edu/radiology/faculty/hasan.html
ExploreDTI	DTI-tracking	http://www.exploredti.com/animations/
FDT TBSS Probtrack	FSL diffusion Tools Voxel-based Prob. tracking	http://www.fmrib.ox.ac.uk/fslcourse/lectures/practicals/fdt/index.htm#tbss http://www.fmrib.ox.ac.uk/fsl/tbss/index.html http://www.fmrib.ox.ac.uk/fslcourse/lectures/fdt.pdf
INRIA MedINRIA SepINRIA	DTI	http://www-sop.inria.fr/asclepios/software/MedINRIA/
Quantitative DTI	DTI	http://www.nitrc.org/projects/quantitativedti/
SATURN	Single Tensor DTI	http://www.lpi.tel.uva.es/saturn
SLICER-DTMRI	DTI	http://www.na-mic.org/Wiki/index.php/Slicer:DTMRI
SPM	fMRI,spatial normalization registration	http://www.fil.ion.ucl.ac.uk/spm/ http://www.fil.ion.ucl.ac.uk/spm/ext/#toolboxes http://users.loni.ucla.edu/∼narr/protocol.php?q=vbm_spm
TrackVis	Qball, DTI, DSI	http://www.trackvis.org/
TracTor	DTI	http://code.google.com/p/tractor/ http://www.nitrc.org/projects/tractor
Volume-One dTV	Diffusion Tensor Visualiser	http://www.volume-one.org/ http://www.ut-radiology.umin.jp/people/masutani/dTV/dTV_frame-e.htm

The first level of DTI quantification uses region-of-interest (ROI) to measure diffusivity and anisotropy (see Figure 6). A second level may use the estimated diffusion tensors, FA or diffusivities for group-wise analyses such as voxel-based analysis (VBA) [115-121], tract-based statistics (TBSS) [13, 122-125] or tensor-based that require multiple subject data alignment and spatial normalization or warping to a template [126-128]. A third level of analysis may use the principal eigenvector combined with anisotropy and multiple ROIs to perform deterministic fiber tracking [12-13, 82-91, 129-130]. A fourth level of analysis may involve using the anisotropy and diffusivity maps to segment the data into white and gray matter which can subsequently be used to perform volume or atlas-based analyses using the international consortium for brain mapping (ICBM http://www.loni.ucla.edu/ICBM/Downloads/Downloads_Atlases.shtml) atlases [131-138]. The regional results of the latter step can be used as masks for probabilistic fiber tracking [139-145]. An extension of the atlas-based approach utilizes the cMRI or T1-weighted data which can be used to generate brain volumes that can be warped unto the DTI data or other data sets such as relaxation, functional MRI, PET and MEG [146- 150].

6.

An illustration of the ROI and fiber tracking in three dimensional space.

3.4 Illustration of Quantitative Diffusion MRI Software Output

We illustrate DTI quantification and ability to cluster different tissue types. In Figure 7, we have used deterministic fiber tracking using multiple ROIs implemented in DTIstudio [95] to track several association, projection and commissural pathways. These pathways were fused with the anatomical cMRI-parcellated results using FreeSurfer. The cMRI and DTI maps were coaligned in SPM and viewed in MRIcro after saving all results as ANALYZE 3D volume files. Figure 8 illustrates the ability to use DTI data to segment the cortical and deep gray matter (e.g. hippocampus, amygdale) and corresponding connections (e.g. uncinate fasciculus) using a DTI atlas-based approach [138].

7.

An illustration of quantitative DTI data obtained using white matter fiber tracking and cortical and regional (e.g. caudate) gray matter segmentation. Several tissue types are represented in the mean diffusivity vs. FA space.

8.

An illustration of DTI atlas-based gray matter segmentation and fusion with fiber tracking and anatomical MRI data after registration.

4. Conclusions

In this work, we have provided an up-to-date review of computational methods applied to diffusion MRI data with focus on the single diffusion tensor model. We have also listed a host of software packages that have been reported in the diffusion MRI literature. To-date there has been no general consensus on the optimal diffusion tensor acquisition protocol (e.g. ∼ 7 minute single tensor vs. ∼ 30 minute DSI). Attention must be paid to data acquisition details (e.g. SNR and spatial resolution) while minimizing image distortion and motion artifacts and keeping the scan time tolerable for imaging children and patients. Quality assurance measures to ensure consistency and accuracy of the DTI data need to be implemented before comparing healthy control and patient data collected for cross-sectional and serial analyses. There are trade-offs, strengths and weaknesses associated with the choice of diffusion data analysis strategy (e.g. deterministic vs. probabilistic tracking, ROI, atlas-based, VBA vs. TBSS).