Elizabeth B Hutchinson1, Alexandru Avram1, Michal Komlosh1, M Okan Irfanoglu1, Alan Barnett1, Evren Ozarslan2, Susan Schwerin3, Kryslaine Radomski3, Sharon Juliano3, and Carlo Pierpaoli1
1SQITS, NICHD/NIH, Bethesda, MD, United States, 2Bogazici University, Istanbul, Turkey, 3APG, USUHS, Bethesda, MD, United States
Synopsis
We have systematically compared four diffusion MRI models –
DTI, DKI, MAP-MRI and NODDI – in the same DWI data sets for fixed brain tissue
to identify the relative strengths of these approaches and characterize the
effects of experimental design and image quality on the generated metrics. Metric-specific advantages in sensitivity and
specificity were shown as well as differential vulnerability across the metrics
to DWI sampling scheme and noise. The
intention of this work is to provide an integrative view of diffusion metrics
that contributes to their utility in brain research.Purpose
New modeling approaches for diffusion MRI data are
promising for improved characterization of non-Gaussian data and may
potentially increase sensitivity and specificity. The first goal of this work was to identify
and evaluate the benefits of new diffusion MRI frameworks that more fully
characterize the diffusion propagator or employ biological
or “microstructure based” modeling. The
second goal of this work was to understand the dependence of metrics on the
experimental design and image quality of acquired DWI data. Both goals address the objective to provide a
systematic evaluation of the next generation of diffusion modeling tools so
that they may be used in the most effective way.
Methods
The general approach of this study was to apply four
different diffusion MRI models using the same input data sets to determine
metric relationships and the effects of image quality and experimental design
on metrics. Comprehensive ex-vivo DWI data sets were acquired for four
mouse brains and one ferret brain at 7T using 3DEPI with isotropic resolution
of 100 and 250 um3 voxel size for mouse and ferret
respectively. The full data set
contained 297 DWIs with 8-shells of b=100-10,000 s/mm2.and was
manipulated to generate two subsampled “experimental design” datasets with: 5-shells
(b=100-1700s/mm2) and 6-shells (b=100-3800s/mm2) and
three additional “image quality” datasets with: 0%, 5%, 10% and 25% added
rectified noise.
Four representative models were applied to the full and manipulated
data sets including:
Diffusion Tensor Imaging (DTI1,2) with metrics of
Trace(D) and fractional anisotropy (FA)
Diffusion kurtosis imaging (DKI3,4) with metrics
of mean kurtosis (Kmean) and kurtosis FA (KFA)
Mean apparent propagator (MAP5) MRI with metrics
of return to the origin probability (RTOP), non-Gaussianity (NG) and propagator
anisotropy (PA)
Neurite orientation dispersion distribution imaging (NODDI6)
with metrics of compartmental volume fractions for intracellular restricted (VIR),
isotropic free (VISO) and intracellular (VIC) as well as
the orientation dispersion index (ODI)
To evaluate between-metric relationships 2D histogram
analysis along with targeted ROI measurements were performed to compare
isotropic and compartmental higher order metrics with TR and to compare higher
order metrics of anisotropy and dispersion with FA.
The effects of experimental design and noise on
each metric were evaluated using within-metric comparisons of whole brain histograms
and maps generated by modeling of subsampled and noise-added data sets.
Results
The Kmean and rtop both demonstrated an inverse
relationship with TR showing high values in
regions of low diffusivity (figures 1 and 2). VIR was close to zero for most gray
matter regions and increased primarily in white matter, while VIC
showed a range of values across tissue types with a large number of voxels
following a negative correlation with TR. It is interesting
to notice that different models
did not show a consistent
monotonic behavior.
For example, the highest VIC
values were in the hypothalamic regions,
while the highest kmean and rtop
values were in the CC, and these
two regions had very similar values of NG.
Comparisons of higher order metrics of anisotropy and
dispersion with FA showed distinct patterns for the whole brain and for regions
of white matter with different fiber geometry (figure 3). The KFA appeared directly related to FA in
white matter, while PA remained high even
in white matter regions that
had low FA because of low intravoxel
orientational coherence, which was nicely revealed
by ODI.
The dependence of the metrics on experimental design varied
widely. TR and FA, as well as VIC
and ODI, showed a shift of the mode of the distribution,
rtop had remarkable stability of the modem, but the shape of the tail at high
values changed, and all other metrics
showed large changes of both histogram
mode and shape (figure 4).
The dependence of metrics on image quality was similarly
variable (figure 5) with the greatest vulnerability found for metrics derived
from the non-Gaussian part of the signal (e.g. Kmean, Kfa, NG, PA) as
well as VIC, and
less vulnerability for other measures (e.g. ODI,
VIR, and rtop) and DTI measures.
Discussion and Conclusion
The results of our study show that higher
order diffusion models can provide discrimination of structural and
architectural features of fixed brain tissue that may not be readily
discriminated on maps of DTI metrics.
However, several metrics
derived from higher order diffusion models show high variability depending on
the experimental design used as well as high noise
susceptibility. As higher order diffusion models will be
gaining more widespread usage, researchers and clinicians should consider
that potential gains in tissue characterization should
be weighted against the intrinsic lower
reliability of these metrics.
Acknowledgements
This work was funded by the Center for Neuroscience and Regenerative Medicine, CDMRP award W81XWH-13-2-0019 and the Henry M. Jackson foundation.References
1. Basser, P.J., Mattiello, J., LeBihan, D., 1994. Estimation of the
effective self-diffusion tensor from the NMR spin echo. J Magn Reson B 103,
247–54.
2.
C. Pierpaoli, L.
Walker, M. O. Irfanoglu, A. Barnett, P. Basser, L-C. Chang, C. Koay, S.
Pajevic, G. Rohde, J. Sarlls, and M. Wu, 2010, TORTOISE: an integrated
software package for processing of diffusion MRI data, ISMRM 18th annual
meeting, Stockholm, Sweden, #1597
3.
Jensen, J.H., Helpern, J.A., Ramani, A.,
Lu, H., Kaczynski, K., 2005. Diffusional kurtosis imaging: the quantification
of non-gaussian water diffusion by means of magnetic resonance imaging. Magn
Reson Med 53, 1432–40.
4.
Tabesh, A., Jensen, J.H., Ardekani, B.A.,
Helpern, J.A., 2011. Estimation of tensors and tensor-derived measures in
diffusional kurtosis imaging. Magn Reson Med 65, 823–36.
5.
Özarslan, E., Koay, C.G., Shepherd, T.M.,
Komlosh, M.E., Irfanoglu, M.O., Pierpaoli, C., Basser, P.J., 2013. Mean
apparent propagator (MAP) MRI: a novel diffusion imaging method for mapping
tissue microstructure. Neuroimage 78, 16–32.
6.
Zhang, H., Schneider, T.,
Wheeler-Kingshott, C.A., Alexander, D.C., 2012. NODDI: practical in vivo
neurite orientation dispersion and density imaging of the human brain.
Neuroimage 61, 1000–16.