Physiological noise modelling for spinal functional magnetic resonance imaging studies
Introduction
The spinal cord is the first site at which sensory information received from the body is processed (Willis and Coggeshall, 2004). Considering the pain system, primary afferent fibres will typically make their first synaptic connection in the dorsal portion of the spinal cord, the “dorsal horn”. Neurons within the dorsal horn are arranged in layers (“laminae”), which broadly correspond to the different types of afferent fibre entering the cord (Willis and Coggeshall, 2004). These connections are subject to top–down and local neuronal influences that will modulate the transmission of nociceptive signals to the brain, and thus have a great impact on pain sensation (Suzuki et al., 2004, Tracey et al., 2002). Indeed, neuropathic pain (Hansson et al., 2001) is thought to arise, at least in part, due to dysfunction (e.g. rearrangement, disinhibition, cell loss) at the primary synaptic connection in the dorsal horn (Woolf and Salter, 2000). Pain-related activity of rat spinal neurons has been assessed using autoradiography (Coghill et al., 1991, Porro and Cavazzuti, 1993), and animal spinal fMRI studies are beginning to be performed (Lawrence et al., 2004, Lilja et al., 2006, Majcher et al., 2006, Malisza and Stroman, 2002, Porszasz et al., 1997). There is a clear need for the development of techniques to record human spinal cord function non-invasively. The close connection between vasculature and neurons within the spinal cord (Giove et al., 2004), should make this structure amenable to the same techniques used to study the brain: e.g. using blood oxygenation level-dependent (BOLD) contrast to infer neuronal activation (Ogawa et al., 1990).
The main difficulties facing researchers attempting to image the spinal cord are its small cross-sectional area, its proximity to structures (vertebrae, intervertebral discs) giving rise to large variations in magnetic susceptibility, the influence of physiological effects that displace the cord or generate signal intensity changes near its superficial layers, and passive displacements of the cord in response to limb movement. Concerning its size, the cord has a maximum diameter at the cervical (neck) level, of approximately 10 mm (Gray, 1918), and is elliptical in shape. Given current limitations on imaging resolution (e.g. with echo-planar imaging (EPI) sequences) at 1.5 T, it is possible to fit at most 25 voxels (with 4 mm2 in-plane resolution) within the cord on axial slice whilst retaining adequate signal-to-noise ratio (SNR), and minimising intra-voxel dephasing (Jones et al., 2001). This places a severe limitation on the ability to resolve individual laminae within the spinal cord, which are at most a millimetre thick. The presence of alternating bone–tissue interfaces along the length of the spinal cord, make adequate shimming extremely difficult (Cooke et al., 2004), and given EPI’s inherent sensitivity to Bo variation this may cause distortion of anatomy in obtained images, as well as global dephasing and concurrent loss of signal. In addition, physiological effects have the ability to alter cord position: these are primarily due to cardiac driven pulsation of cerebrospinal fluid (CSF) within the subarachnoid space surrounding the cord (Greitz et al., 1993). Flow of unsaturated CSF “spins” into the imaging slice will give rise to temporal variation in the CSF signal adjacent to the cord, which through partial volume effects, and finite point spread function, may give rise to noise within the voxels that always contain spinal cord. Respiratory effects may either take the form of bulk movement of spinal and connective tissue due to shifting tissue mass as the chest wall rises/falls, or bulk susceptibility effects as previously noted in the brain (Windischberger et al., 2002), which is of greater significance to functional imaging of the spinal cord or brainstem. Lastly, and of particular concern to studies employing motor tasks, is the fact that limb movements give rise to millimetre displacements in the position of the nerve root(s) serving the moved limb (Miyamoto et al., 2003), which might translate to small (stimulus correlated) movements of the cord, but may also, in the case of neck flexion, give rise to large longitudinal shifts of the entire cord (Kuwazawa et al., 2006).
Physiological effects on CSF flow have been investigated previously using flow-sensitive EPI (e.g. Friese et al., 2004b, Klose et al., 2000). These studies demonstrated that it was possible to observe cardiac and respiratory effects in critically sampled (short TR ∼ 160 ms) single-slice EPI data taken through either the fourth ventricle or at various positions along the spinal cord. CSF was found to flow caudally during systole and rostrally during diastole (demonstrating its cardiac dependence). However, the magnitude of signal changes (proportional to flow rate) occurring as fresh CSF spins flowed into the imaging slice, also depended on the position in the respiratory cycle when the image was acquired. Interaction of cardiac and respiratory processes has previously been observed in the spectra of temporally reordered brain EPI data (Frank et al., 2001). The presence of these physiological effects suggests possible avenues for reducing physiologically derived signal noise. In particular, the k-space and image–space approaches of Hu et al. (1995) and Glover et al. (2000), respectively. Both techniques assign cardiac and respiratory phases to each acquired image slice based on their acquisition time relative to a separate physiological recording (e.g. ECG, pulse oximeter, respiratory bellows). These are then modelled through a low order Fourier expansion to remove their effect from recorded images. The image space technique (RETROICOR) has also been implemented via a general linear model (GLM) approach (Corfield et al., 1999, Restom et al., 2006).
To date, several groups have reported functional imaging data for the cervical and lumbar spinal cord (Backes et al., 2001, Govers et al., 2007, Komisaruk et al., 2002, Madi et al., 2001, Stroman et al., 2002b, Stroman et al., 2002c), using a variety of different stimulation paradigms. Studies using spin-echo-based imaging (Stroman, 2005, Stroman et al., 2001a, Stroman et al., 2001b, Stroman et al., 2002a, Stroman and Ryner, 2001) have consistently demonstrated spinal cord activity. Data from gradient-echo-based acquisition schemes, however, have proved to be largely inconsistent. Explanations for the differences between these results may relate to e.g. the absence of any physiological noise correction scheme; the use of cardiac gating for acquisition of functional images, which will introduce T1-dependent signal changes in recorded images, that are unrelated to the task; the use of breath holding techniques to remove or reduce respiratory-related motion, which will likely produce hypercapnic effects (Hoge et al., 1999) mimicking BOLD signal changes; or the failure to use any motion correction when analysing recorded images.
We have previously demonstrated the utility of RETROICOR in aiding the identification of spinal cord activation (Brooks et al., 2006) when using gradient-echo-based imaging. Similar approaches have been applied to modelling the effect of physiological noise on brain images, e.g. Lund et al. (2006) who demonstrated that the noise structure in EPI time series may be partially whitened by including the RETROICOR components into the GLM. A recent report has addressed the issue of physiological noise modelling in spin-echo-based spinal functional imaging (Stroman, 2006), by decomposing the aliased recorded physiological information (pulse and respiration) into its principal components, and including them as nuisance regressors in the GLM. In this paper, we sought to investigate sources of physiological noise affecting gradient-echo-based acquisitions, to examine if this technique can provide complimentary information to the spin-echo method.
Increasingly, imaging of the brainstem is being performed to investigate its role in pro- and anti-nociceptive processes (Dunckley et al., 2005, Fairhurst et al., 2007, Zambreanu et al., 2005), and this structure will suffer from similar physiological effects that affect the spinal cord. For this reason, and the others stated above, an adequate physiological noise model (PNM) is required. To determine the various types and sources of physiological noise present in spinal cord images, data were acquired under two conditions: (1) during rest; and (2) during painful thermal stimulation. We acquired resting images with both short (TR = 200 ms “critically sampled”: 5 Hz) and long (TR = 3 s “under-sampled”: 0.33 Hz) TRs. Exploratory data analysis (using probabilistic independent component analysis (PICA) (Beckmann and Smith, 2004)) was used to investigate the various physiological noise components present. The aim was to characterise the sources of physiological noise present in spinal imaging, to inform the design of a new PNM – a modification of the conventional RETROICOR approach. Various PNMs were then tested using F-tests to examine whether the model explained a significant amount of residual signal variance (physiologically related noise) above that which could be attributed to over-fitting the noise, using the same spinal images. Finally the candidate PNMs were included in the GLM to record spinal activation in response to painful thermal stimulation, and thereby assess the improvement in modelled signal that may be obtained using an optimised PNM.
For clarity, this paper is split into three sections. The acquisition and basic pre-processing of images was common to both experiments utilising resting data, so is described first, followed by separate sections including Methods, Results and Discussion relevant to: (1) physiological signal identification based on resting data (using PICA) for the exploration of possible sources of variability; (2) model-based analysis of resting data (within the framework of a GLM) to determine the improvement in noise modelling when using a PNM; and (3) the application of the optimal PNM to activation data (spinal cord functional imaging). Lastly, a general discussion of the analysis techniques used, and of the spinal fMRI data, is presented.
Section snippets
Methods
Eight healthy subjects (4 female) aged between 25 and 46 years, were imaged with a 1.5 T Siemens Sonata MR system (Siemens Medical Systems, Erlangen, Germany). Subjects were placed supine on the scanner bed, and their neck supported by a custom built Perspex former. A dual-element phased array surface coil was placed between the subjects’ neck and the support, and suitable padding placed beneath the head. No further restraint was used to minimise movement. To monitor heart and respiration rate,
Discussion
By applying model-free analysis (MELODIC) to critically sampled (short TR) spinal EPI data, we were able to clearly distinguish cardiac and respiratory effects as well as interactions between these signals. Components of similar spatial distribution were identified using the same approach with a long TR acquisition at the sampling rate used to record multi-slice functional images of the spinal cord. The presence of these physiologically driven components motivated the modification to the
Conclusions
We have demonstrated that, using exploratory data analysis, it is possible to inform the design of a modified PNM (based on RETROICOR) for modelling physiological noise in spinal EPI data. The choice of components to include in the PNM was made based on F-tests, which revealed that RETRO37 was the most effective model. This was also confirmed in the analysis of activation data. Combining the power of the PNM with a model-based GLM provides an automated method that incorporates physiological
Acknowledgments
The authors would like to acknowledge the financial support of the Human Frontiers Science Program RGP 0013/2004 (JCWB, CAP, IT), BBSRC (MJ), HEFCE (IT), EPSRC (CFB), RAE/EPSRC (KLM) and MRC (RGW). The authors would also like to thank Dr Matthew Robson and Dr Daniel Gallichan, for their help in modifying pulse sequences for spinal imaging.
References (59)
- et al.
Separating respiratory-variation-related fluctuations from neuronal-activity-related fluctuations in fMRI
NeuroImage
(2006) - et al.
Anticipatory brainstem activity predicts neural processing of pain in humans
Pain
(2007) - et al.
Thresholding of statistical maps in functional neuroimaging using the false discovery rate
NeuroImage
(2002) - et al.
Issues about the fMRI of the human spinal cord
Magn. Reson. Imaging
(2004) - et al.
Improved optimization for the robust and accurate linear registration and motion correction of brain images
NeuroImage
(2002) - et al.
Correlation of functional activation in the rat spinal cord with neuronal activation detected by immunohistochemistry
NeuroImage
(2004) - et al.
Non-white noise in fMRI: does modelling have an impact?
NeuroImage
(2006) - et al.
Simultaneous functional magnetic resonance imaging in the rat spinal cord and brain
Exp. Neurol.
(2006) - et al.
Spatial and temporal aspects of spinal cord and brainstem activation in the formalin pain model
Prog. Neurobiol.
(1993) - et al.
Physiological noise reduction for arterial spin labeling functional MRI
NeuroImage
(2006)