Tianjia Zhu1,2, Minhui Ouyang1, Nikou Lei3, David Wolk4, Paul Yushkevich5, and Hao Huang1,5
1Department of Radiology, Children's Hospital of Philadelphia, Philadelphia, PA, United States, 2Department of Bioengineering, University of Pennsylvania, Philadelphia, PA, United States, 3Department of Physics, University of Washington Seattle, Seattle, WA, United States, 4Department of Neurology, University of Pennsylvania, Philadelphia, PA, United States, 5Department of Radiology, University of Pennsylvania, Philadelphia, PA, United States
Synopsis
Diffusion
MRI (dMRI) has ushered in a new era in which conventional brain cortical histological
measures such as soma and neurite densities may be assessed noninvasively
through advanced dMRI. However, analytical dMRI microstructural models are restricted
by the model assumptions and lack of validation from quantitative histology
data. Individual dMRI parameters characterize only limited microstructural
information. By leveraging a variety of dMRI-based parameters delineating
cortical microstructure from multiple aspects, we established a machine
learning based method accurately estimating cortical soma and neurite densities
in the cortex, paving the way for data-driven noninvasive virtual histology for
potential applications to Alzheimer’s diseases.
Purpose
Diffusion
MRI (dMRI) has ushered in a new era in which conventional brain cortical histological
measures such as soma and neurite densities may be assessed noninvasively
through advanced dMRI1-9. However, analytical dMRI microstructural models
are restricted by the model assumptions and lack of validation from quantitative
histology data. Individual dMRI parameters characterize only limited
microstructural information. Besides DTI-based metrics which are sensitive to brain
microstructures10,11, metrics derived from diffusion kurtosis
imaging (DKI) are sensitive to neurofilament density in the cerebral cortex9.
Here, by leveraging a variety of high resolution dMRI-based parameters
delineating cortical microstructure from multiple aspects, we established a
machine learning based method accurately estimating cortical soma and neurite
densities in the cortex.Methods
The
method overview is shown in Fig.1. DMRI (Fig 1 left panel): High
resolution dMRI acquisitions with two b-values (b=1500s/mm2, b=4500s/mm2)
and 30 independent gradient directions were performed on a normal postmortem
Rhesus macaque brain. DMRI coronal slices were oriented to be consistent with coronal
slices in histology12 (brainmaps.org). DMRI parameters were: In-plane
resolution: 0.6×0.6mm FOV=100×100×72mm, in plane imaging matrix=166×166, slice
thickness=2mm, TR/TE=2100/77.8ms, NSA=24, 2 repetitions for each b-value. DTI
and DKI fitting (Fig 1 left panel): After correction current distortion of
eddy and affine registration of diffusion weighted images, diffusion tensor and
kurtosis were fitted. Axial (AK),
radial (RK) and mean kurtosis (MK) as well as fractional anisotropy (FA), axial
(AD), radial (RD) and mean diffusivity (MD) (Fig 3A) maps were obtained.
Quantification of soma and neurite densities from histological images: As
shown in Fig 2A, for measuring soma density (SD), the Nissl-stained histology
images of resolution 0.46 µm/pixel were blocked into segments of 512 × 512
pixels with each segment size of 0.24 × 0.24 mm. Each segment is converted to
gray scale and threshold. SD is defined as number of contoured areas/ segment
area in mm2. The calculated SD map agrees well with soma density
from histology studies13. As shown in Fig 2B, for measuring neurite
density (ND) based on histological image, a structure tensor was computed for
every pixel in the histological image14. Pixels with Anisotropy Index
(AI) > 0.6 were classified as fiber structure. The neurofilament stained
images were blocked into segments of 512 × 512 pixels with each segment size of
0.24 × 0.24 mm. In each blocked segment,
the ratio of the area classified as fiber structure to the blocked area was defined
ND. Histology-MRI registration and
cortical segmentation:
Neurofilament histology slice was converted to grayscale and affine registered
to b0. ND map was registered to b0 using the same transformation.SD map was
directly affine registered to b0. Cortical areas were segmented based on
intensity and overlapping regions on b0 and histology. The segmentation results
are shown in Fig.3B. Random forest (RF) regression to estimate soma and
neurite densities: RF regression was implemented in Scikit-learn15.
The cortical masks were applied to AK, RK, MK, FA, AD, RD, MD, SD, and ND maps
and each voxel in the cortical mask served as a training sample. A RF regressor
with max-depth=7, number of estimators=1000, max-number of features=5, and mean
squared error (MSE) loss was used to predict SD. For predicting ND, max-depth=7,
number of estimators=1000, max-number of features=3 and MSE was used. Model
performance was evaluated using five-fold cross-validation, and average feature
importance were extracted.Results
Fig 4 shows estimation results in one validation set. Both the estimated
SD (Fig 4A) and the ND(Fig 4B) correlate significantly with
ground-truth values. (P<0.0001, Pearson correlation coefficient r=0.53 and
r=0.53 respectively). Across five folds, the average mean absolute error (MAE)
for predicting soma density is 0.02×104 somas/mm2, with
average r=0.46 for correlation between predicted and ground truth soma density.
The MAE for predicting neurite density is 0.04, with average r=0.49 for the correlation.
Fig. 4C shows the feature importance. DKI-derived MK ranks the first in feature
importance in estimating both soma and neurite densities. For estimating ND,
AD, AK, RK also have high importance.Discussion and Conclusion
We have built a machine
learning based method to infer cortical soma and neurite densities directly
from dMRI measures. To the
best of our knowledge, this is the first study estimating cortical histological
measures using machine learning of a variety of dMRI-derived measurements, filling the gap between MRI contrasts
and cell densities quantified from histology. DMRI-derived metric
maps (Fig 3) demonstrate high resolution and high SNR to delineate
microstructure of the cerebral cortex, despite relatively smaller size of
macaque brains compared to that of human brains. DKI measures in general achieved higher feature importance than DTI
measures, consistent with previous studies9. Furthermore, highest
importance of MK in both SD and ND predictions shows it as a good candidate for
inputs to more advanced modeling. The relatively high feature importance of
axial and radial measures (AD, AK, RK) for predicting ND is consistent with the
fibrous structure of neurites and serves as a feature-engineering step for
future, more advanced analytical and machine learning models. This study can be
further improved by adopting better MRI-histology registration and larger
datasets. Our study
demonstrates the possibility for potentially paradigm-shifting data-driven
virtual histology in Alzheimer’s disease with dMRI.Acknowledgements
This study is funded by NIH
MH092535, MH092535-S1 and HD086984.References
1.
Stanisz, G. J., Wright, G. A., Henkelman, R.
M., & Szafer, A. (1997). An analytical model of restricted diffusion in
bovine optic nerve. Magnetic Resonance in Medicine, 37(1),
103-111.
2.
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. Magnetic
Resonance in Medicine: An Official Journal of the International Society for
Magnetic Resonance in Medicine, 53(6), 1432-1440.
3.
Zhou, X. J., Gao, Q., Abdullah, O., &
Magin, R. L. (2010). Studies of anomalous diffusion in the human brain using
fractional order calculus. Magnetic resonance in medicine, 63(3),
562-569.
4.
Jespersen, S. N., Kroenke, C. D., Østergaard,
L., Ackerman, J. J., & Yablonskiy, D. A. (2007). Modeling dendrite density
from magnetic resonance diffusion measurements. Neuroimage, 34(4),
1473-1486.
5.
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(4),
1000-1016.
6.
Le Bihan, D., & Johansen-Berg, H. (2012).
Diffusion MRI at 25: exploring brain tissue structure and function. Neuroimage, 61(2),
324-341.
7.
Alexander, D. C., Dyrby, T. B., Nilsson, M.,
& Zhang, H. (2019). Imaging brain microstructure with diffusion MRI:
practicality and applications. NMR in Biomedicine, 32(4),
e3841.
8.
Novikov, D. S., Fieremans, E., Jespersen, S.
N., & Kiselev, V. G. (2019). Quantifying brain microstructure with
diffusion MRI: Theory and parameter estimation. NMR in Biomedicine, 32(4),
e3998.
9.
Zhu, T., Peng, Q., Ouyang, A., & Huang, H.
(2020). Neuroanatomical underpinning of diffusion kurtosis measurements in the
cerebral cortex of healthy macaque brains. Magnetic Resonance in
Medicine.
10. Song, S.
K., Sun, S. W., Ju, W. K., Lin, S. J., Cross, A. H., & Neufeld, A. H.
(2003). Diffusion tensor imaging detects and differentiates axon and myelin
degeneration in mouse optic nerve after retinal ischemia. Neuroimage, 20(3),
1714-1722.
11. Song, S.
K., Sun, S. W., Ramsbottom, M. J., Chang, C., Russell, J., & Cross, A. H.
(2002). Dysmyelination revealed through MRI as increased radial (but unchanged
axial) diffusion of water. Neuroimage, 17(3),
1429-1436.
12. Mikula,
S., Trotts, I., Stone, J. M., & Jones, E. G. (2007). Internet-enabled
high-resolution brain mapping and virtual microscopy. Neuroimage, 35(1),
9-15.
13. Carlo, C.
N., & Stevens, C. F. (2013). Structural uniformity of neocortex,
revisited. Proceedings of the National Academy of Sciences, 110(4),
1488-1493.
14. Budde, M.
D., & Frank, J. A. (2012). Examining brain microstructure using structure
tensor analysis of histological sections. Neuroimage, 63(1),
1-10.
15. Pedregosa et al Scikit-learn: Machine
Learning in Python., JMLR 12, pp. 2825-2830, 2011.