IMPACT OF REALIGNMENT IN SPINAL fMRI Julien COHEN-ADAD 1,2 , Mathieu PICHÉ 1 , Pierre RAINVILLE 1 , Habib BENALI 2 and Serge ROSSIGNOL 1 1 GRSNC, Department of Physiology, Faculty of Medicine, Université de Montréal, Montreal, QC, Canada (e-mail: julien.cohen-adad@imed.jussieu.fr ). 2 INSERM U678, Université Pierre et Marie Curie (Paris VI), CHU Pitié-Salpêtrière, Paris, France. Context What is fMRI? Functional magnetic resonance imaging (fMRI) allows for the detection of task-related neuronal activity through acquisition of successive MRI volumes. Analysis of functional MRI series mostly relies on a voxel-by-voxel comparison of signals with and without performing a task (e.g., motor task, visual stimuli). What is realignment? Given the duration of the data acquisition, subjects might move in the scanner and the location of every voxel might vary between scans. To correct for these motion artifacts, a preprocessing of fMRI time series is usually applied for spatial realignment. It consists in an inter- scan rigid body co-registration based on voxel intensity matching [1]. Cardiac noise is prominent in spinal fMRI Why is it problematic? Since cardiac noise mostly contributes to variance in spinal data, one could expect different outcomes of realignment procedure when performed in spinal versus brain fMRI time series. Indeed, realignment estimation is based on a majority of voxels that have the same rigid motion. Since the effect of cardiac pulsations is widespread in spinal imaging and thus induces in-phase variation of a large number of voxels, estimation of the motion correction matrix might include these variations. Thus, the objective of this paper is to investigate the impact of realignment on spinal fMRI time series. Methods MRI acquisition Acquisition was performed on healthy volunteers (N=4) using a 1.5T Siemens Avanto MR scanner. For each subject, four runs were acquired. Each run consisted of a 6 minute acquisition over the cervical cord (single shot spin-echo EPI, sagittal orientation, TR/TE = 250/40 ms, voxel size = 2×2×2 mm). The low TR was chosen to avoid any aliasing effect of the cardiac signal. External cardiac signal was recorded using a plethysmograph. MRI processing Raw data were realigned using SPM software [1]. Both raw and realigned data were then band-pass filtered respectively to the external cardiac recording, to recover most of the cardiac variance. Data were then averaged within each run. A mask including the spinal canal was used to quantify cardiac variance between raw and realigned data. Figure 2 summarizes the various processing steps. Results Correlation between cardiac recording and registration parameters Realignment procedure did have a significant impact on the distribution of cardiac variance within spinal fMRI time series. Following our hypothesis, we observed high coherence (i.e., correlation in frequency) between some realignment parameters and external cardiac recording. Figure 3 illustrates such a coherence for one subject, with associated coherence values for each realignment parameter after having removed low frequency drifts. Reduction of cardiac variance The proposed method allowed us to quantify the cardiac variance within a region of interest containing the cervical spinal canal (i.e., spinal cord with cerebrospinal fluid, from C2 to T1). Realigned data showed reduced cardiac variance through the whole volume (64 ± 8 %), as well as within the spinal canal (39 ± 3 %) as shown in Figure 5. Sagittal topography of cardiac variance before and after realignment is provided in Figure 4. Discussion Realignment might spread focal activation When haemodynamic-related activation is spatially narrow, realignment procedure might smooth activated voxels, thus reducing the sensitivity of activation detection. By analogy, a large activated area might lead to spurious activation in other part of the imaged volume, as already reported [4]. Impact of GLM analysis on realigned data When analyzing fMRI data using the General Linear Model (GLM), an assumption is made regarding the independent and identically normally distributed (i.i.d.) characteristic of residuals. In the case of spinal fMRI data, we would expect high temporal autocorrelations within residuals due to cardiac fluctuations (if not modeled), thus leading to a biased T-score [5]. We showed that data realignment might drastically modify cardiac contribution within fMRI time series, thus modifying GLM results. How to limit cardiac influence when realigning? We suggest either (i) to use of a mask prior to realignment in order to estimate parameters of rigid transformation based on voxels without most of the cardiac activity (e.g., skin interface), or (ii) to reduce cardiac variance prior to realignment using various detrending methods. Figure 3. Parameters of the transformation matrix (left, top and middle) and cardiac recording (left, bottom) for one subject. Pearson coefficient and associated T-score for each transformation parameters compared to cardiac recording in the frequency domain (top table). 66.84 0.95 RZ* 64.23 0.94 RY* 44.39 0.89 RX* 48.02 0.91 TZ* 13.18 0.51 TY* 0.27 0.01 TX T-score r Transfo 40 50 60 70 80 90 100 -1 -0.5 0 0.5 time (s) mm x translation y translation z translation 40 50 60 70 80 90 100 -5 0 5 10 x 10 -3 time (s) deg x rotation y rotation z rotation 40 50 60 70 80 90 100 -5 0 5 time (s) a.u. cardiac recording V i,j Vr i,j Vf i,j Vrf i,j vf i vrf i F i m i filter average mask cardiac recording realign vf im vrf im Figure 2. Processing steps for run i, volume j. V i,j and Vr i,j are respectively the raw and realigned data. F i is the band-pass filter, m i is the designed mask specific to subject and run. References [1] K.J. Friston, S. Williams, R. Howard, R. S. Frackowiak, and R. Turner, "Movement-related effects effects in fMRI time-series," Magn Reson Med, vol. 35, pp. 346-55, Mar 1996. [2] F. Giove, G. Garreffa, G. Giulietti, S. Mangia, C. Colonnese, and B. Maraviglia, "Issues about the fMRI of the human spinal cord," Magn Reson Imaging, vol. 22, pp. 1505-16, Dec 2004. [3] M. Piché, J. Cohen-Adad, M. Kosh Nejad, V. Perlbarg, G. Xie, G. Beaudoin, H. Benali and P. Rainville, “Characterization of cardiac-related noise in fMRI of the cervical spinal cord,” Neuroimage (submitted). [4] L. Freire and J.F. Mangin, "Motion Correction Algorithms May Create Spurious Brain Activations in the Absence of Subject Motion," NeuroImage, vol. 14, pp. 709-722, 2001. [5] T. Lund, K. Madsen, K. Sidaros, W.-L. Luo, and T. Nichols, "Non-white noise in fMRI: Does modelling have an impact?," NeuroImage, vol. 29, pp. 54-66, 2006. Acknowledgments We thank G. Beaudoin and the staff of the CHUM Notre-Dame for their support in data acquisition. This work was supported by the Canada Research Chair on the Spinal Cord provided by the Canadian Institute of Health Research (CIHR) for a fellowship to J.C-A and by the Multisdisciplinary Team on Locomotor Rehabilitation (Regenerative Medicine and Nanomedicine, CIHR). BOLD signal in spinal cord has been reported to range from 0-5% [2] while cardiac-related signal may vary up to 50-90% with high inter- subject variability [3]. Cardiac signal arises from cerebrospinal fluid pulsation and from the main vessels surrounding the spinal cord. As an example, Figure 1 provides a topography of cardiac variance. 1 2 3 4 500 1000 1500 2000 2500 3000 Subject Cardiac variance (a.u.) raw realigned Figure 5. Cardiac variance before and after realignment, averaged within spinal canal. For each subject, results are averaged across four runs, with standard deviation indicated by error bars. A two- tailed Student’s test indicated a significant diminution of cardiac variance within the region of interest (*P<0.005, **P<0.0005). ** * ** * Figure 1. Cardiac variance in the spinal canal (in % to the baseline). 48 10 Figure 4. Sagittal topography of cardiac variance before (top) and after (bottom) realignment for all four subjects (left to right). The color code represents MRI intensity values (in % to the baseline). 9 94 8 75 7 72 9 55