Denoising spinal cord fMRI data: Approaches to acquisition and analysis
Introduction
Functional magnetic resonance imaging (fMRI) of the spinal cord is still at a relatively early stage. While the first reports of spinal fMRI appeared in the late 1990s (human imaging: Stroman et al., 1999; Yoshizawa et al., 1996; animal imaging: Pórszász et al., 1997), the field is still relatively small – there are currently about 100 published reports in humans and less than 50 in animals on spinal imaging with fMRI. In this review we will solely focus on the human spinal imaging literature, although exciting advances have been made with regard to spinal imaging in rats (e.g. Lawrence et al., 2004; Lilja et al., 2006; Zhao et al., 2009), cats (Cohen-Adad et al., 2009a), and very recently also in monkeys (Chen et al., 2015, Yang et al., 2015), where fine-grained sensory processing patterns were investigated as well as changes in resting-state connectivity due to spinal cord injury. It will be exciting to see how these approaches develop, especially when considering their application in tandem with other techniques such as two-photon imaging (Johanssen and Helmchen, 2013) or optogenetics (Montgomery et al., 2016).
A likely reason for the relative dearth of spinal fMRI studies lies in the difficulty of obtaining reliable results from this structure, which is due to a number of factors (Giove et al., 2004, Stroman, 2005, Stroman et al., 2014, Summers et al., 2010). First of all, the spinal cord has a very small cross-sectional diameter (approximately 12mm in the left-right and 8mm in the anterior-posterior direction at the cervical level; Fradet et al., 2014; Ko et al., 2004), which makes it necessary to use a high in-plane resolution to minimize partial-volume effects. Second, the presence of tissue-types with different magnetic susceptibility (vertebrae and connective tissue) leads to inhomogeneities in the static magnetic field (B0) that result in a periodic signal disruption along the superior-inferior axis of the spinal cord (Cooke et al., 2004, Finsterbusch et al., 2012). Third, the influence of physiological noise (arising from respiratory and cardiovascular sources; Fig. 1) on blood-oxygen-level-dependent (BOLD) responses is much stronger in the spinal cord than in the brain (Cohen-Adad et al., 2010, Piché et al., 2009), with the potential for obscuring task-related responses and resting-state connectivity profiles, unless adequately addressed.
With regards to the sources of physiological noise, one can dissociate respiratory and cardiac influences (though interactions are also observed; Brooks et al., 2008). Respiratory influences are mostly evident in time-dependent distortions of the B0 field time-locked to the respiratory cycle (Raj et al., 2001) – these manifest as apparent non-rigid motion of the spinal cord and are much stronger than in the brain due to the spinal cord's proximity to the lungs (Verma and Cohen-Adad, 2014). Cardiac influences manifest themselves in vascular pulsations (Dagli et al., 1999) on the one hand, which might obscure spinal cord BOLD responses since the main arteries and veins run directly along the edge of the cord in close proximity to the minute grey matter regions of interest (Cohen-Adad et al., 2010). On the other hand, cardiac activity leads to pulsatile movement of the cerebrospinal fluid (CSF; O’Connell, 1943) in which the spinal cord is embedded, resulting in a non-rigid oscillatory cord motion time-locked to the cardiac cycle (Figley and Stroman, 2007). Cardiac induced CSF flow furthermore leads to large signal variations due to unsaturated spins moving into an imaging slice (Finsterbusch et al., 2012), which is highly problematic considering the proximity of the CSF-containing subarachnoid space to the spinal cord grey matter.
In spite of all these difficulties, there is great interest in studying the human spinal cord non-invasively, since it is the brain's principal sensorimotor interface with the body – thus properly characterising input to and output from the central nervous system relies on knowledge of spinal cord activity. Furthermore, there is also strong clinical interest in the spinal cord, due to its prominent involvement in a number of neurological disorders (Wheeler-Kingshott et al., 2014). For example, in multiple sclerosis, deficits in motor function may be due to dysfunction at multiple levels of the neuro-axis, and early spinal cord involvement in the primary progressive form of the disease has recently been demonstrated (Abdel-Aziz et al., 2015). Equally, damage occurring to motor neurons in amyotrophic lateral sclerosis have been shown to involve both the cortico-spinal tract and the spinal cord (Foerster et al., 2013). In patients with spinal cord injury the motivation to study the cord is clear and current interest includes relating cord atrophy to disability and monitoring treatment success (Cadotte and Fehlings, 2013, Freund et al., 2013), though imaging presents unique challenges due to the use of metal in stabilising the vertebral column. Lastly, the development and maintenance of chronic pain states in animal models has been shown to depend on altered spinal cord function (Woolf and Salter, 2000) and spinal cord fMRI would allow researchers to examine whether similar changes underlie the transition to chronic pain in man.
Here we set out to give an overview of approaches that have been used to overcome some of the aforementioned problems. We will first look at how one can optimise the signal-to-noise ratio (SNR) during image acquisition, based on a variety of methods not routinely used in brain imaging. Then we will move on to describe the analysis approaches that have been developed to deal with the various sources of noise present in spinal fMRI data – here we will also mention differences when denoising task-based versus resting-state data. We will briefly mention various motion correction and spatial filtering approaches before concentrating on the main topic of denoising using either modelling approaches based on externally acquired signals (cardiac and respiratory data) or data-driven approaches that aim to find structured noise in the data that is then used in clean-up procedures. We will conclude with an outlook on procedures that have proven helpful for denoising brain fMRI data and might hold potential for spinal fMRI as well.
Section snippets
Acquisition approaches to denoising
Before coming to acquisition methods that optimize the SNR in the spinal cord, it is important to briefly review the gross anatomy and vascular supply of the spinal cord (Baron, 2015, Silverdale, 2015, Thron, 2016). From the superior to inferior (rostral to caudal) direction, the spinal cord is organized into cervical, thoracic, lumbar, sacral, and cogzygeal parts; hereafter we will focus on studies of the cervical spinal cord as this part has received the greatest attention. Each of these
Data inspection and motion correction
A crucial first step to analysing spinal fMRI data is visual inspection of the recorded time series data – this might highlight problems such as large movements (e.g. induced by swallowing) or increases in background noise (e.g. ghosting) that affect discrete volumes. In these situations, there are several approaches to deal with the affected volumes, such as replacing the affected volumes by the average of its neighbours, replacing the affected volumes by the time-series average, using
Outlook
In closing, we would like to highlight a few approaches that have shown beneficial effects for noise reduction in the brain. Since many of the problems that these approaches aim to ameliorate exist in the spinal cord to a larger degree, their application to spinal fMRI might be of significant benefit.
References (114)
- et al.
Groupwise multi-atlas segmentation of the spinal cord's internal structure
Med. Image Anal.
(2014) - et al.
Reproducibility of resting state spinal cord networks in healthy volunteers at 7 T
NeuroImage
(2016) - et al.
Data-driven optimization and evaluation of 2D EPI and 3D PRESTO for BOLD fMRI at 7 T: I. Focal coverage
NeuroImage
(2011) Modelling with independent components
NeuroImage
(2012)- et al.
A component based noise correction method (CompCor) for BOLD and perfusion based fMRI
NeuroImage
(2007) - et al.
Advances in functional magnetic resonance imaging of the human brainstem
NeuroImage
(2014) - et al.
Assessment of data acquisition parameters, and analysis techniques for noise reduction in spinal cord fMRI data
Magn. Reson. Imaging
(2014) - et al.
Spinal cord functional MRI at 3 T: gradient echo echo-planar imaging versus turbo spin echo
NeuroImage
(2008) - et al.
Improved sensitivity and specificity for resting state and task fMRI with multiband multi-echo EPI compared to multi-echo EPI at 7 T
NeuroImage
(2015) Physiological noise modeling and analysis for spinal cord fMRI
Physiological noise modelling for spinal functional magnetic resonance imaging studies
NeuroImage
BOLD fractional contribution to resting-state functional connectivity above 0.1 Hz
NeuroImage
BOLD signal responses to controlled hypercapnia in human spinal cord
NeuroImage
Investigations on spinal cord fMRI of cats under ketamine
NeuroImage
Array coils
Localization of cardiac-induced signal change in fMRI
NeuroImage
An adaptive filter for suppression of cardiac and respiratory noise in MRI time series data
NeuroImage
Development and validation of retrospective spinal cord motion time-course estimates (RESPITE) for spin-echo spinal fMRI: Improved sensitivity and specificity by means of a motion-compensating general linear model analysis
NeuroImage
B0 inhomogeneity and shimming
Functional neuroimaging of inner fields-of-view with 2d-selective RF excitations
Magn. Reson. Imaging
Single, slice-specific z-shim gradient pulses improve T2*-weighted imaging of the spinal cord
NeuroImage
Issues about the fMRI of the human spinal cord
Magn. Reson. Imaging
The nuisance of nuisance regression: spectral misspecification in a common approach to resting-state fMRI preprocessing reintroduces noise and obscures functional connectivity
NeuroImage
Origin of subarachnoid cerebrospinal fluid pulsations: a phase-contrast MR analysis
Magn. Reson. Imaging
The impact of physiological noise correction on fMRI at 7T
NeuroImage
Improved optimization for the robust and accurate linear registration and motion correction of brain images
NeuroImage
Integration of motion correction and physiological noise regression in fMRI
NeuroImage
Visual inspection of independent components: defining a procedure for artifact removal from fMRI data
J. Neurosci. Methods
Assessment of physiological noise modelling methods for functional imaging of the spinal cord
NeuroImage
Differentiating BOLD and non-BOLD signals in fMRI time series using multi-echo EPI
NeuroImage
Correlation of functional activation in the rat spinal cord with neuronal activation detected by immunohistochemistry
NeuroImage
Resting-state fMRI confounds and cleanup
NeuroImage
Reduction of motion-related artifacts in resting state fMRI using aCompCor
NeuroImage
Spectral characteristics of resting state networks
Prog. Brain Res.
CORSICA: correction of structured noise in fMRI by automatic identification of ICA components
Magn. Reson. Imaging
Characterization of cardiac-related noise in fMRI of the cervical spinal cord
Magn. Reson. Imaging
Recent progress and outstanding issues in motion correction in resting state fMRI
NeuroImage
Evaluation of ICA-AROMA and alternative strategies for motion artifact removal in resting state fMRI
NeuroImage
ICA-AROMA: A robust ICA-based strategy for removing motion artifacts from fMRI data
NeuroImage
Automatic denoising of functional MRI data: combining independent component analysis and hierarchical fusion of classifiers
NeuroImage
Susceptibility artefacts
Spatio-temporal point-spread function of fMRI signal in human gray matter at 7 T
NeuroImage
The current state-of-the-art of spinal cord imaging: methods
NeuroImage
Functional exploration of the human spinal cord during voluntary movement and somatosensory stimulation
Magn. Reson. Imaging
Comparison of physiological noise at 1.5 T, 3 T and 7 T and optimization of fMRI acquisition parameters
NeuroImage
Evidence for early neurodegeneration in the cervical cord of patients with primary progressive multiple sclerosis
Brain
Combined diffusion weighting and CSF suppression in functional MRI
NMR Biomed.
Resting state functional connectivity in the human spinal cord
eLife
Simultaneous multislice (SMS) imaging techniques
Magn. Reson. Med.
Cited by (47)
Advanced spinal cord MRI in multiple sclerosis: Current techniques and future directions
2022, NeuroImage: ClinicalCitation Excerpt :Another study at 3T observed a link between increased resting-state connectivity in the CSC sensory network in pwRRMS and greater tissue damage as measured by DTI, suggesting that functional compensatory mechanisms may exist in the cord similarly to those observed in the brain in the early disease stages (Combes et al., 2022). Further clinical applications will be facilitated as advances in acquisition (Barry et al., 2018a; Kinany et al., 2022), quality assessment and pre-processing (Eippert et al., 2017) are made. The use of 7T can benefit SC fMRI via increased sensitivity to susceptibility effects for BOLD effect-based (T2*) contrast, and greater SNR which enables acquisitions with sub-millimeter in-plane resolution.
Investigating the human spinal sensorimotor pathways through functional magnetic resonance imaging
2021, NeuroImageCitation Excerpt :In terms of anatomical image pre-processing, quasi-automatic segmentation methods for the WM, GM and CSF, including surface-based, intensity-based or image-based methods, have also been recently developed, thus permitting a precise characterization of the anatomical structure of the spinal cord in both healthy individuals and patients with spinal pathologies (De Leener et al., 2016). Second, the development of new fMRI acquisition protocols at high and ultra-high field strengths, shimming procedures, optimized pulse sequences and selective field-of-view techniques, combined with more sophisticated functional preprocessing methods such slice-wise correction (De Leener et al., 2017) and physiological noise modeling (Brooks et al., 2008; Eippert et al., 2017a) have dramatically increased the reliability of spinal fMRI findings. Finally, the development of a standard template of the spinal cord (e.g., PAM50; De Leener et al., 2018) provides now the possibility to average results across subjects while reducing selection and registration biases, and then perform group-based analyses as well as compare groups of participants in the same reference space, similar to standard procedures used for brain fMRI data.