Zhengshi Yang1, Xiaowei Zhuang1, Karthik Sreenivasan1, Virendra Mishra1, and Dietmar Cordes1,2
1Cleveland Clinic Lou Ruvo Center for Brain Health, Las Vegas, NV, United States, 2Department of Psychology and Neuroscience, University of Colorado, Boulder, CO, United States
Synopsis
The fluctuation introduced by head motion, cardiac and respiratory
fluctuations and other noise sources considerably confounds the interpretation
of resting-state fMRI data. These noise fluctuations widely spread the whole
brain regardless of the kinds of brain tissues, however, neural activity is
more likely limited to gray matter tissue. Considering that the contribution of
neural activity varies in different brain tissues, we hypothesized that disentangling
gray matter and non-gray matter time series can clean fMRI data and improve the
data quality. With such a hypothesis, we proposed a deep neural network method
to denoise resting state fMRI data.
Introduction
The fluctuation introduced by head motion, cardiac and respiratory
fluctuations and other noise sources considerably confounds the interpretation
of resting-state fMRI data [1]. The complex
mechanism between the noise source and their contributed fluctuation in fMRI
data makes it challenging to clean fMRI data. These noise fluctuations widely
spread the whole brain regardless of the kinds of brain tissues, however,
neural activity is more likely limited to gray matter tissue. Considering that
the contribution of neural activity varies in different brain tissues, we
hypothesized that disentangling gray matter (GM) and non-gray matter (non-GM) time
series can clean fMRI data and improve the data quality. With such a
hypothesis, we proposed a deep neural network method to denoise resting state
fMRI data.Methods
The structural MRI and resting state fMRI (rsfMRI) data used in this
study are publicly available in ADNI database (http://adni.loni.usc.edu/). Both T1 and rsfMRI
data were normalized to MNI template space. rsfMRI timeseries were divided to
two categories, namely GM timeseries and non-GM (white matter and ventricle)
time series, based on segmented T1 images. Each GM voxel timeseries is randomly
paired with one non-GM voxel timeseries and each pair of time series is treated
as one sample to optimize the deep neural network. The neural network is
trained on each subject separately and generates a subject-specific model.
There are about 50,000 samples for each subject, which is enough to train our
neural network. The neural network consists of seven layers as shown in Fig.1a.
The first layer is a time-dependent fully-connected layer, which has a
fully-connected layer for each time point but with different parameters. This
layer can be treated as a deep learning based scrubbing technique to remove
spike artifacts. However, unlike scrubbing removes time points and requires
arbitrary hard threshold, such a layer keeps all time points since they can be
informative and learns from the data to properly separate signals from
artifacts. Then a concatenation layer is used to remove the dummy (2nd)
dimension, followed by two temporal convolutional layers with 32 and 16
filters, respectively. The temporal convolutional layers play a role as
low-pass filtering but without a fixed frequency threshold. Finally three
time-distributed fully-connected layers, which have a fully-connected layer
with the same parameters for all time points, are used and output the denoised
timeseries. A schematic plot of fully-connected layer, time-distributed
fully-connected layer and time-dependent fully-connected layer is shown in
Fig.1b. The neural network is optimized by minimizing the correlation between
GM timeseries and non-GM timeseries in each pair for the purpose of
disentangling timeseries between tissues. Along with this denoising neural
network (DeNN), nuisance regression was also performed for comparison. The
strategies used in each denoised time series were listed in Fig.2.Result
To examine the specificity [2] of each
denoising methods, we compared the connectivity values of four regions within
default mode network with the posterior cingulate cortex (PCC) seed (-7, -55,
27). The proposed DeNN network has the highest specificity for all these four
regions, including left lateral parietal cortex (LLP), medial prefrontal cortex
(MPFC), PCC and right lateral parietal cortex (RLP) (see Fig.3). In addition,
MPFC (-1, 49, -2) was shown to be functionally unrelated to two visual
reference regions (10-mm spheres around (-30, -88, 0) and (30, 88, 0)) [3]. As shown in Fig.4, we observed that orig, 12P, 24P
timeseries have significantly positive bias between MPFC and visual reference
regions, and both 14P and14P+GS have significantly negative bias. In contrast,
DeNN denoised time series has weakest correlation between MPFC and bilateral visual
reference regions. Furthermore, voxelwise fractional amplitude of low frequency
fluctuation (fALFF) [4] and voxelwise framewise
displacement (FDvox) [5] were calculated for all
the subjects. The group-level voxelwise correlation between fALFF and FDvox was
calculated and shown in Fig.5. The correlation map for original data suggested
that FDvox overall is anti-correlated with fALFF. Such a relation remains for
12P and 24P data, and is more severe for 14P and 14P+GS. In contrast, the
anti-correlation between FDvox and fALFF is alleviated in 14P+aCompcor, and is
even weaker for DeNN. For DeNN denoised data, there is no voxel passing the
correlation threshold 0.2. Discussion
In this study, we proposed a deep neural network method to denoise
resting state fMRI data by disentangling timeseries between different brain
tissues. Instead of generating a set of nuisance regressors as in traditional
denoising methods, this method directly outputs the denoised timeseries and
remains the degrees of freedom in the original data. Compared with other
methods, DeNN method does not have either positive or negative bias with
reference regions, has the highest the specificity for the regions within
default mode network, and has the best performance in alleviating the
correlation between fALFF and quality control measurement. Conclusion
A robust and automated deep neural network framework is proposed to
reduce noise fluctuation from multiple noise sources by disentangling
timeseries between brain tissues.Acknowledgements
This research project was supported by the NIH (grant 1R01EB014284 and
COBRE grant 5P20GM109025), Young Investigator Award from Cleveland Clinic, a private grant from Peter and Angela Dal Pezzo, a private grant from Lynn and William Weidner, and a private grant from Stacie and Chuck Matthewson. Data
collection and sharing for this project was funded by the Alzheimer's Disease
Neuroimaging Initiative (ADNI) (National Institutes of Health Grant U01
AG024904) and DOD ADNI (Department of Defense award number W81XWH-12-2-0012).References
[1]
Caballero-Gaudes, C., Reynolds, R.C., 2017. Methods for cleaning the BOLD fMRI
signal. Neuroimage 154, 128-149.
[2]
Weissenbacher, A., Kasess, C., Gerstl, F., Lanzenberger, R., Moser, E.,
Windischberger, C., 2009. Correlations and anticorrelations in resting-state
functional connectivity MRI: a quantitative comparison of preprocessing
strategies. Neuroimage 47, 1408–1416.
[3]
Van Dijk, K.R., Hedden, T., Venkataraman, A., Evans, K.C., Lazar, S.W.,
Buckner, R.L., 2010. Intrinsic functional connectivity as a tool for human
connectomics: theory, properties, and optimization. J. Neurophysiol. 103,
297–321.
[4]
Q.H. Zou, C.Z. Zhu, Y. Yang, X.N. Zuo, X.Y. Long, Q.J. Cao, Y.F. Wang, Y.F.
Zang. An improved approach to detection of amplitude of low-frequency
fluctuation (ALFF) for resting-state fMRI: fractional ALFF J. Neurosci.
Methods, 172 (2008), pp. 137-141, 10.1016/j.jneumeth.2008.04.012
[5]
Yan, C., et al, 2013. A comprehensive assessment of regional variation in the
impact of head micromovements on functional connectomics. NeuroImage 76
183-201.