Assessment of physiological noise modelling methods for functional imaging of the spinal cord

Abstract

The spinal cord is the main pathway for information between the central and the peripheral nervous systems. Non-invasive functional MRI offers the possibility of studying spinal cord function and central sensitisation processes. However, imaging neural activity in the spinal cord is more difficult than in the brain. A significant challenge when dealing with such data is the influence of physiological noise (primarily cardiac and respiratory), and currently there is no standard approach to account for these effects. We have previously studied the various sources of physiological noise for spinal cord fMRI at 1.5 T and proposed a physiological noise model (PNM) (Brooks et al., 2008). An alternative de-noising strategy, selective averaging filter (SAF), was proposed by Deckers et al. (2006). In this study we reviewed and implemented published physiological noise correction methods at higher field (3 T) and aimed to find the optimal models for gradient-echo-based BOLD acquisitions. Two general techniques were compared: physiological noise model (PNM) and selective averaging filter (SAF), along with regressors designed to account for specific signal compartments and physiological processes: cerebrospinal fluid (CSF), motion correction (MC) parameters, heart rate (HR), respiration volume per time (RVT), and the associated cardiac and respiratory response functions. Functional responses were recorded from the cervical spinal cord of 18 healthy subjects in response to noxious thermal and non-noxious punctate stimulation. The various combinations of models and regressors were compared in three ways: the model fit residuals, regression model F-tests and the number of activated voxels. The PNM was found to outperform SAF in all three tests. Furthermore, inclusion of the CSF regressor was crucial as it explained a significant amount of signal variance in the cord and increased the number of active cord voxels. Whilst HR, RVT and MC explained additional signal (noise) variance, they were also found (in particular HR and RVT) to have a negative impact on the parameter estimates (of interest) — as they may be correlated with task conditions e.g. noxious thermal stimuli. Convolution with previously published cardiac and respiratory impulse response functions was not found to be beneficial. The other novel aspect of current study is the investigation of the influence of pre-whitening together with PNM regressors on spinal fMRI data. Pre-whitening was found to reduce non-white noise, which was not accounted for by physiological noise correction, and decrease false positive detection rates.

Introduction

The spinal cord is the first relay site in the transmission of sensory information from the periphery to the brain (D'Mello and Dickenson, 2008, Willis and Coggeshall, 2003). To allow a more complete understanding of central nervous system processing in health and disease, a non-invasive technique for assessing the function of human spinal cord is desirable. To address this need, functional magnetic resonance imaging (fMRI) of the spinal cord has been developed. Since the publication of the first article on spinal cord fMRI in 1996 (Yoshizawa et al., 1996), there have been more than 50 papers on spinal cord fMRI performed in humans and animals. Spinal cord activation has been observed in cervical and lumbar spinal cord, for stimuli such as touch, vibration, thermal, brush and motor tasks (for review, Giove et al., 2004, Leitch et al., 2010, Stroman, 2005b).

Despite the increasing numbers of spinal cord fMRI studies in the literature, the characteristics of the functional signal is not fully understood (Bouwman et al., 2008, Cohen-Adad et al., 2010, Stroman, 2005a), and reported human data have been inconsistent. Possible reasons for this include: small cord area (~ 1 cm²) producing intrinsically low signal to noise data; magnetic susceptibility differences (at bone/disc interfaces) causing signal loss and image distortion; physiological noise (primarily cardiac and respiratory related) obscuring functional signals; motion and flow artefacts due to cerebrospinal fluid (CSF) pulsation; variability of the results across repeated measurements and lack of a standard coordinate template for group analysis (Leitch et al., 2010, Stroman et al., 2008). One of the key challenges remaining to establish spinal cord fMRI as a routinely applied technique is to adequately account for physiological noise effects when analysing data.

The blood oxygenation level dependent (BOLD) signal in fMRI is sensitive to a range of physiological variables. Physiological noise can induce changes in the BOLD signal and can therefore obscure the detection of neural activation using traditional experimental paradigms, and the relative contribution to measured signal could vary with increasing field strength, different acquisition parameters or multi-channel array coils (Bodurka et al., 2007, Hutton et al., 2011, Kruger and Glover, 2001, Triantafyllou et al., 2005, Triantafyllou et al., 2011). Physiological noise is a significant confound in connectivity studies, particularly estimation of the resting state (Birn et al., 2008a, Chang and Glover, 2009a, Chang and Glover, 2009b, Jo et al., 2010, Lund, 2001, Shmueli et al., 2007). Several physiological noise sources have been studied in the brain fMRI literature, including the dominant low frequency fluctuations from the subject motion, cardiac and respiratory processes (Bhattacharyya and Lowe, 2004, Bianciardi et al., 2009, Birn et al., 2006, Chang et al., 2009, Gray et al., 2009a, Lund et al., 2006, Shmueli et al., 2007, Wise et al., 2004). Various methods of physiological noise correction for brain fMRI have been proposed e.g. k-space correction (Hu et al., 1995), independent component analysis (ICA) denoising (Kelly et al., 2010, Tohka et al., 2008) or image-based correction (Glover et al., 2000), known as RETROspective Image CORrection (RETROICOR). RETROICOR models cardiac and respiratory induced signal changes using a Fourier basis series defined by the relative phases in each cycle (i.e. cardiac and respiratory) at the time of image acquisition. More recently Deckers et al. (2006) proposed a selective averaging filter (SAF) method to account for the cardiac and respiratory effects by averaging imaging data based on their acquisition time relative to the cardiac and cycles. Heart rate (HR) (Chang et al., 2009, Shmueli et al., 2007) and respiratory related fluctuations (respiration volume per time — RVT, Birn et al., 2006) have been found to explain additional variance of the BOLD signal recorded in the brain.

Functional images of the brainstem and spinal cord are more severely affected by physiological noise (Brooks et al., 2008, Diedrichsen et al., 2010, Harvey et al., 2008) than the brain (Cohen-Adad et al., 2010). In particular, their proximity to the throat/trachea/lungs produces B0-induced susceptibility effects that can lead to signal change and image movement (Pattinson et al., 2009, Raj et al., 2001, Van de Moortele et al., 2002). The relative proximity of these structures (and relative sizes) to the major arteries lying near their surface and the CSF-filled spaces surrounding them (Anderson et al., 2009, Naidich et al., 2009), means that cardiac noise is particularly problematic (Dagli et al., 1999). The pulsatile nature of arterial blood and CSF flow around the cervical spinal cord, leads to large signal fluctuations near the boundaries of these structures (Piché et al., 2009). Indeed, by using hypercapnia as a positive control stimulus, Cohen-Adad et al. found that the highest responses at the spinal level were detected outside the cord and may reflect the contribution from larger veins or CSF pulsation effects (Cohen-Adad et al., 2010).

An improved RETROICOR method has been applied to brainstem fMRI, and gave a significant improvement in the detection of task-related activation (Harvey et al., 2008). The complex spatiotemporal structure of cardiac noise in the cervical spinal cord was recently examined, and ICA found to be useful when correcting for the cardiac effects (Piché et al., 2009). The use of principal component analysis (PCA) to decompose recorded cardiac parameters into their constituent signals has been proposed, and shown to account for their presence in spinal functional data (Figley and Stroman, 2009, Stroman, 2006). However this proposed PCA method has been optimised mainly for cardiac-related noise with spin-echo acquisition parameters, as in these images the respiratory effects are not as significant compared to data acquired with gradient-echo readouts. The SAF method was recently applied to a placebo analgesia spinal cord fMRI study (Eippert et al., 2009), but the performance of this correction scheme was not discussed. We have previously characterised various sources of physiological noise affecting spinal cord imaging and demonstrated the utility of a physiological noise model (PNM) for fMRI (Brooks et al., 2008). By including in the General Linear Model (GLM), basis functions derived from physiological measurements of cardiac and respiratory processes, and their interaction and a regressor defined by the measured CSF signal intensity time course, false-positive detection rates of pain-related activity recorded at 1.5 T were reduced and the expected location of activation was revealed.

The advantage of model-based techniques is that they are derived from actual physiological measurements from each subject, and the associated physiological noise can be automatically removed by including the model as nuisance regressors in the GLM. The noise structure in EPI time series could therefore be partially whitened by including these nuisance regressors into the GLM (Lund et al., 2006). Pre-whitening of fMRI data is advantageous for accurate statistical inference using the GLM, as temporal noise in the time series is assumed to be random (white); if it is not, then the statistical inference will be less accurate (Smith et al., 2007, Woolrich et al., 2001). A previous study compared the performance of the GLM when using RETROICOR regressors with and without an AR(1) model, and the AR(1) model was shown to be inferior to physiological noise modelling when used as a pre-whitening step (Lund et al., 2006).

In this study different model-based physiological noise correction techniques for spinal cord fMRI will be reviewed and implemented under the framework of the GLM. We evaluated the performance of different physiological noise models by comparing the residuals after model fitting, F-test regression model comparisons and the number of activated voxels under different types of stimulation. We aimed to find the optimal physiological noise model for spinal cord gradient-echo-based BOLD acquisitions, and explored whether pre-whitening performed using FMRIB's Improved Linear Model (FILM, Woolrich et al., 2001) is a necessary step to control false positive detection rates. Functional responses in the spinal cord of the rat have previously been studied using autoradiography with thermal pain stimuli (Coghill et al., 1991) and fMRI with electrical stimuli (Lilja et al., 2006). Thermal pain and sensory stimuli have been applied in human spinal cord fMRI studies, and segmental responses observed following stimulation of different dermatomes (Ghazni et al., 2010, Lawrence et al., 2008, Stroman, 2009). In the present study, thermal pain and punctate stimuli were used to investigate spinal cord responses. A summary of abbreviations used in this paper is given in Table 4.