Reduction of physiological noise with independent component analysis improves the detection of nociceptive responses with fMRI of the human spinal cord

Abstract

The evaluation of spinal cord neuronal activity in humans with functional magnetic resonance imaging (fMRI) is technically challenging. Major difficulties arise from cardiac and respiratory movement artifacts that constitute significant sources of noise. In this paper we assessed the Correction of Structured noise using spatial Independent Component Analysis (CORSICA). FMRI data of the cervical spinal cord were acquired in 14 healthy subjects using gradient-echo EPI. Nociceptive electrical stimuli were applied to the thumb. Additional data with short TR (250 ms, to prevent aliasing) were acquired to generate a spatial map of physiological noise derived from Independent Component Analysis (ICA). Physiological noise was subsequently removed from the long-TR data after selecting independent components based on the generated noise map. Stimulus-evoked responses were analyzed using the general linear model, with and without CORSICA and with a regressor generated from the cerebrospinal fluid region. Results showed higher sensitivity to detect stimulus-related activation in the targeted dorsal segment of the cord after CORSICA. Furthermore, fewer voxels showed stimulus-related signal changes in the CSF and outside the spinal region, suggesting an increase in specificity. ICA can be used to effectively reduce physiological noise in spinal cord fMRI time series.

Introduction

Following the success of functional magnetic resonance imaging (fMRI) in the investigation of brain function, fMRI of the spinal cord was shown to be technically feasible in both humans (Backes et al., 2001, Bouwman et al., 2008, Brooks et al., 2008, Cohen-Adad et al., 2010, Eippert et al., 2009, Giulietti et al., 2008, Govers et al., 2007, Madi et al., 2001, Maieron et al., 2007, Majcher et al., 2007, Stracke et al., 2005, Stroman et al., 2004, Summers et al., 2010, Valsasina et al., 2008, Yoshizawa et al., 1996) and animals (Cohen-Adad et al., 2009, Endo et al., 2008, Lawrence et al., 2007, Lilja et al., 2006, Majcher et al., 2007, Zhao et al., 2008). However, results remain controversial due to the low reproducibility of spinal cord fMRI (Bouwman et al., 2008, Giove et al., 2004). Because of the relatively small cross-sectional size of the spinal cord grey matter, even very little motion can contaminate the signal. Spinal cord motion and in-flow effects from the surrounding cerebrospinal fluid (CSF) along with cardiac and respiratory cycles greatly degrade the quality of fMRI data by adding unwanted variance in the time series (Figley and Stroman, 2007, Giove et al., 2004, Kong et al., 2012, Stroman, 2006, Stroman et al., 2005). Non-invasive acquisition of functional images of the spinal cord is very challenging but remains much needed even though there is currently no methodological consensus on the adequate means to address this challenge.

Several methods exist to minimize physiological-related noise. Respiratory-gated acquisition and breath-hold have been employed to reduce the effect of respiratory activity (Stroman and Ryner, 2001, Stroman et al., 1999), although with moderate success (Stroman, 2005). Cardiac gating has also been used in the brain (Guimaraes et al., 1998) and spinal cord (Backes et al., 2001) but this requires substantially longer acquisition time. Moreover, variable TR can introduce additional signal variance due to T₁-effects, which are difficult to correct in the spinal cord due to the necessity of acquiring a robust T₁ map. Additionally, heart rate can correlate with the experimental paradigm — especially for painful stimuli, therefore spurious activations can appear in the statistical map as TR would also correlate with heart rate (Tousignant-Laflamme et al., 2005).

Post-hoc correction can be used to reduce physiological noise in fMRI time series, as shown in the brain (Behzadi et al., 2007, Deckers et al., 2006, Glover et al., 2000, Hu et al., 2005, Lund et al., 2006, Thomas et al., 2002, Tohka et al., 2008) and spinal cord (Brooks et al., 2008, Figley and Stroman, 2007, Kong et al., 2012, Stroman, 2006). One method based on the RETROspective Image CORrection (RETROICOR) algorithm (Glover et al., 2000) consists in estimating a set of regressors based on external physiological recordings (pulse oxymeter and respiration trace) to be included in the general linear model (GLM). Although it has shown great success in capturing the variance of physiological noise in several fMRI studies, this approach might inappropriately model the shape of physiological-related signal by a combination of sine and cosine functions. Alternatively, data-driven methods aim at extracting the physiological noise part of the fMRI data from the data itself. One such method performs CORrection of structured noise using Spatial Independent Component Analysis (CORSICA) (Perlbarg et al., 2007). The principle of CORSICA is to estimate a noise map via the spatial independent component analysis (ICA) and to remove the corresponding components from the functional data before testing for the effect of interest. The noise components can be identified using anatomical priors, as in (Perlbarg et al., 2007), or based on separate data acquisition using a short TR to assess the spatial distribution of noise at rest. Short-TR data are used to avoid aliasing, particularly for cardiac signal, whose main spectrum typically ranges from 0.8 to 1.4 Hz. Therefore a sampling frequency of at least 2 × 1.4 Hz (or TR < 350 ms) is required to satisfy the Nyquist condition. CORSICA proved to be useful in reducing physiological noise in brain fMRI time series (Schrouff et al., 2011, Vanhaudenhuyse et al., 2010). The main assumption of CORSICA is that physiological noise is spatially structured. Our recent investigation on the spatial distribution of physiological noise in spinal cord fMRI demonstrated that cardiac-related noise — one major source of signal variance in spinal cord fMRI — is spatially structured and stable within each individual (Piché et al., 2009). Brooks et al. performed ICA on spinal cord fMRI data acquired with short TR (200 ms) to allow exploration of the physiological noise spectra (Brooks et al., 2008). They showed that cardiac and respiratory effects are easily extracted by the ICA decomposition. The cardiac-related signal was mostly located in the CSF region and in the carotid and vertebral vessels, while the respiratory-related signal usually appeared at the interface between connective tissues (neck, muscles). The interaction between cardiac and respiratory signal (amplitude modulation) also appeared in the ICA. In addition, a low-frequency component (< 0.1 Hz) was robustly identified in multiple subjects. All these observations suggest that a map of physiological noise can be computed for each subject and subsequently used within the same subject in various experiments.

In this paper we assessed whether CORSICA can improve the sensitivity and specificity of BOLD responses to nociceptive stimuli in the cervical spinal cord.