Suheyla Cetin Karayumak1,2, Marek Kubicki1,2, and Yogesh Rathi1,2
1Harvard Medical School, Boston, MA, United States, 2Brigham and Women's Hospital, Boston, MA, United States
Synopsis
Diffusion
MRI (dMRI) data obtained from a 7T scanner has novel and improved
microstructural tissue information missing from data acquired on 3T scanners. In
this work, we propose to use deep Convolutional Neural Networks (CNN) that use
rotation invariant spherical harmonic (RISH) features to map the dMRI data (the
raw signal) between scanners without changing the fiber orientation. We validate
our algorithm on 40 Human Connectome Project (HCP) subjects with scans on both 3T
and 7T (10 training + 30 test). Our preliminary results on 30 test subjects
shows that CNN can indeed reliably obtain 7T dMRI data quality from 3T scans.
Introduction
In recent years, enormous efforts have been made
to improve the quality of MRI. Higher strength MR scanners (7T) have been
employed to provide better contrast and resolution in diffusion-weighted (DW)
imaging1,2,3,4. However, very few 7T scanners are currently
available for research purposes. Recently, several machine learning methods
have been proposed to boost the resolution and quality of 3T scanners5,6,
or harmonize various dMRI sites7. However, none of these works have
demonstrated their ability to obtain 7T-like dMRI data quality from 3T data. In
this work, we propose to learn an efficient mapping of multi-shell dMRI signal
with different spatial resolution and magnetic field strength (from 3T to 7T)
using RISH features.
Methods
RISH Features: We represent the dMRI signal $$$\mathbf{S}$$$ in a basis of spherical harmonics (SH):
$$$\mathbf{S}\approx\sum_{l}\sum_{m}C_{lm}Y_{lm}$$$ and construct five RISH
features which can be scaled to modify the dMRI signal without changing the
principal diffusion directions of the fibers7:
$$\left|C_l\right|^2=\sum_{m=1}^{2m+1}C_{lm}^2~~~~~~~~~ [1]$$ for $$$l=\left\{0, 2,4,6,8\right\}$$$
(See Figure 1).
Patch-based non-linear mapping of RISH
features from 3T to 7T using CNNs:
Using 3T RISH features as input, our goal is to learn a nonlinear mapping of 3T
to 7T as a patch-wise regression problem. Such mapping can be learned using the
paired 3T and 7T RISH features of training data. To match the spatial
resolution, we first upsample 3T DW volumes using 7th order B-spline
interpolation. Then, we align 3T and 7T data as follows: First, we register b0
maps of 3T and 7T data through rigid registration8. The estimated
transformation is then applied to each DW volume. Next, the gradient vectors
are rotated using the rotation matrix estimated through rigid registration.
After 3T and 7T DW data are aligned, we compute RISH features as in Eq. 1.
We utilize a CNN with five convolutional layers to learn the
non-linear mapping from 3T to 7T. Specifically, we used an $$$n\times n$$$ RISH
feature patch to learn the mapping, where $$$n=7$$$ was used in this work.
Figure 2 summarizes our deep CNN architecture. In each layer, RISH features are
convolved with a $$$3 \times 3$$$ kernel with the number of $$$32$$$, $$$64$$$,
$$$128$$$, $$$256$$$ and $$$256$$$ convolutional filters respectively. After each
convolution step, ReLU operation is applied.
Results
We used 10 HCP subjects as training subjects and another
unseen 30 HCP subjects as test subjects with dMRI scans obtained from both 7T
($$$1.05\times1.05\times1.05mm^3$$$, two-shells: $$$b=\left\{1000, 2000\right\}$$$ and $$$65$$$ gradient directions) and 3T
($$$1.25\times1.25\times1.25mm^3$$$, three-shells: $$$b=\left\{1000, 2000, 3000\right\}$$$ and $$$90$$$ gradient directions) scanners9. We learned
the mapping for $$$b=1000$$$ and $$$b=2000$$$ shells from 3T to 7T.
Subject-specific mapping between 3T and 7T was obtained and the final signal
was estimated using the scaled SH coefficients. We computed
Fractional-Anisotropy (FA) and Mean-Diffusivity (MD) measures to compare the
learning performance between our method (CNN-RISH) and RF-DTI5 which aims to
learn DTI features using Regression Forests. Mean Squared Error (MSE) was
computed for FA and MD between 7T data and our results in test subjects.
Average accuracy and precision values in the estimation of FA, MD,
generalized-FA (GFA) and DWI are given in Table 1. In Figure 3 top row, we
depict the estimated FA results for $$$b=1000$$$ (Fig.3a) and $$$b=2000$$$
(Fig.3b). In the bottom row, we show the MSE maps in FA between the predicted
data and the actual scanner acquired 7T data.
Even though FA and MD are directly related to DTI, our method
performed better when compared to RF-DTI. As seen in Table 1 and in Figure 3,
our method gives the best performance with lowest error in several metrics (FA,
GFA and dMRI signal error). Thus, our method is tissue model-independent and
directly reconstructs the dMRI signal, which can then be used in further
analysis.Discussion and Conclusion
In this paper, we present a CNN based non-linear regression
method which learns the mapping of RISH features and consequently the dMRI
signal from 3T to 7T scanner. Our validation results on 30 test subjects shows
the efficacy of using this technique to gain new information about tissue
structure present in the 7T data but not in 3T data. Further, the proposed
approach is model-free and can be used to harmonize the raw dMRI data from 3T
and 7T scanners. Both qualitative and quantitative results show that our method
performs better when compared to RF-IQT. We note that this work is preliminary
and extensive validation on different datasets is required to further
understand the power and limitations of this technique.
Acknowledgements
The authors would like to acknowledge the following grants which supported this work: R01MH102377, (PI: Dr. Marek Kubicki), R01MH097979 (PI: Dr. Yogesh Rathi).References
-
Chilla GS, Tan CH, Xu C, Poh CL. Diffusion weighted magnetic resonance imaging and its recent trend—a survey. Quantitative Imaging in Medicine and Surgery. 2015;5(3):407-422.doi:10.3978/j.issn.2223-4292.2015.03.01.
- Sotiropoulos SN, Hernández-Fernández M, Vu T, Andersson JL, et al. Fusion in diffusion MRI for improved fibre orientation estimation: An application to the 3T and 7T data of the Human Connectome Project. NeuroImage, 2016;134:396-409.
- Polders DL, Leemans A, Hendrikse J, Donahue MJ, Luijten PR, Hoogduin JM. Signal to noise ratio and uncertainty in
diffusion tensor imaging at 1.5, 3.0, and 7.0 Tesla. J. Magn. Reson.
Imaging. 2011;33: 1456-1463. doi:10.1002/jmri.22554.
-
Vu A, Auerbach E, Lenglet C, et al. High resolution whole brain diffusion imaging at 7 T for the Human Connectome Project. NeuroImage. 2015;122:318-331. doi:10.1016/j.neuroimage.2015.08.004.
- Alexander DC, Zikic D, Ghosh A, Tanno R, Wottschel V, Zhang J, Kaden E, Dyrby TB, Sotiropoulos SN, Zhang H, Criminisi A. Image quality transfer and applications in diffusion MRI. NeuroImage. 2017;152: 283-298. doi:10.1016/j.neuroimage.2017.02.089.
- Tanno R, Worrall DE, Ghosh A, et al. Bayesian Image QualityTransfer with CNNs: Exploring Uncertainty in dMRISuper-Resolution. MICCAI. 2017;611–619.
- Mirzaalian, H., Ning, L., Savadjiev, P. et al. Brain Imaging and Behavior (2017). https://doi.org/10.1007/s11682-016-9670-y.
-
Avants BB, Tustison NJ, Song G, Cook
PA, Klein A, Gee JC. A Reproducible Evaluation of ANTs Similarity Metric
Performance in Brain Image Registration. NeuroImage. 2011;54(3):2033-2044. doi:10.1016/j.neuroimage.2010.09.025.
-
Van Essen DC, Smith SM, Barch DM, et al. The WU-Minn Human Connectome Project: An Overview. NeuroImage. 2013;80:62-79. doi:10.1016/j.neuroimage.2013.05.041.