Slice acquisition order and BOLD frequency content: an eventrelated functional magnetic resonance imaging study. A-L Paradis *, # , P-F Van de Moortele *, D Le Bihan *, J-B Poline *

* Service Hospitalier Frédéric Joliot, CEA, Orsay, France #

Laboratoire de Physiologie de la Perception et de l'Action, CNRS – Collège de

France, Paris, France Keywords: event-related fMRI, acquisition, statistical analysis Abstract: Many event-related fMRI paradigms performed so far have been designed to study a limited part of the brain with high temporal resolution. However, event-related paradigms can be exploratory therefore requiring whole brain scans and so repetition times (TR) of several seconds. For these large TR values, the slice acquisition order may have an important effect on the detection of event-related activation. Indeed, when the scanning is interleaved, the temporal delay between the acquisition of two contiguous slices can reach a few seconds. During this time, the subject is likely to move, and the haemodynamic response (HR) will vary significantly. In this case, the interpolation applied between contiguous slices for motion correction induces a temporal smoothing between voxels that are spatially close but temporally sampled a few seconds apart. This should modify the frequency structure of the response and may impair the detection of short events. We therefore tested the effect of three acquisition schemes (sequential, sequential with gap and interleaved) at two repetition times (TR = 3 sec and TR = 6 sec on 6 and 7 subjects respectively) on activation detection and frequency content in a visual motion event-related paradigm. Unexpectedly, for large TR (6s), results were found in favour of the interleaved acquisition scheme (p < 0.05). For smaller TR, no strong bias could be found. Generally, intra-subject variability (across acquisition schemes) is found much smaller than inter-subject variability, confirming the importance of multi-subjects analyses. Our study also shows that important physiological information is carried by high frequency components that should not be filtered out.

1

INTRODUCTION Event-related fMRI has recently emerged as a powerful approach to study many sensorimotor or cognitive paradigms. An event-related fMRI experiment records the individual haemodynamic responses induced by very short stimuli (from several hundred milliseconds to a few seconds). This approach has several important advantages over the more commonly performed block designed paradigms. In particular, it is free of habituation or learning effects, it can take advantage of the subject score or performance on an individual basis (e.g. error on trial, see (1)) and it generally allows for purer cognitive or sensorimotor components to be studied. Many event-related fMRI paradigms performed so far have been designed to study a limited and small part of the brain with a high temporal resolution. For instance Buckner et al. (2) used a repetition time (TR) of 1 sec, Schad et al. (3) and Wiener et al. (4) used TRs of 80 ms and 320 ms, while Josephs et al. (5) used a TR of 1.7 sec. However, eventrelated paradigms are also likely to be exploratory, and, as such, they might require whole brain scanning. The whole brain constraint may imply TR values up to 5-6 sec (for example in (6), TR = 6.4 sec), depending on the number of slices required and the capabilities of the MRI system. Generally, the detection of some brain BOLD responses to very short stimuli or individual events can be difficult. First, very brief stimuli may lead to a weaker BOLD response and second, the frequency structure of the event-related response is modified as compared to block paradigm responses. In particular, it contains relatively more information in the high frequencies. Consequently, acquisition should be optimised to preserve as much the frequency content of the event-related responses as possible. Many MRI acquisition sequences are based on interleaved slice acquisition, known to yield generally optimal signal to noise ratio (7). However, contiguous slices are acquired at quite different times (with TR = 6 sec, the time between contiguous slices is 6/2 = 3 sec). Thus, movements of the subject in and out of the slice acquisition plane will not only cause spatial shifts of the cerebral structures but also a substantial temporal missampling of the MR signal when realigning images. Furthermore, movement correction implies a spatial interpolation that induces some temporal smoothing of the voxel time-courses. This temporal smoothing is likely to have a small or negligible effect if contiguous slices are acquired at brief intervals, but may be of importance in the interleaved mode for the large TR values used in whole brain experiments. In particular, this may affect the high frequencies of the event-related blood oxygenation level dependent (BOLD) response. A simple simulation study was conducted and showed that the energy of the HR signal could be decreased by 20% (TR= 3sec) to 50% (TR=6sec) when the simulated movement along the z-axis was of the order of half a voxel. However, a realistic simulation of all processes (acquisition, sampling, spin history, movement and realignment) was not found tractable. Therefore, only an experimental study could inform in a reliable manner on this question. The natural alternative to interleaved acquisition is sequential slice acquisition. It is well known that sequential slice acquisition may cause loss of MR signal due to rapid and successive spin excitation in the region where the small side lobes of the excitation slice

2

profile overlap (7). However, it may be that this loss of signal is small compared to the temporal missampling and interpolation effects. To avoid slice profile overlap, a small gap (e.g. 1mm) is generally added between slices. The latter acquisition mode can be thought of as a trade off between the sequential and interleaved schemes, since the subject movements during the time required to acquire one single slice are expected to be small (less than the gap size). In a recent work, Howseman et al. (8) performed similar acquisition schemes. Their study addressed the effect of in-flow and slice thickness on BOLD activation signal with a block design paradigm in a single subject experiment, and they report little or no in-flow effect. In this study, we aimed at evaluating the influence of the slice acquisition scheme on the detectability of an event related activation signal. We therefore performed experiments to evaluate the effect of three slice acquisition modes on the statistical results obtained with an event-related fMRI activation experiment. In particular, we aimed at determining whether loss of signal due to the temporal smoothing expected in a real dataset acquired with an interleaved acquisition scheme would be greater or not than the loss of signal expected in a sequential acquisition scheme due to cross-talk. We also tested the frequency content of the BOLD signal to investigate the modifications of the activation signal depending on the acquisition scheme. This paper presents for the first time results from an experiment specifically designed to study the effect of the slice acquisition scheme on the statistical results of an event-related fMRI paradigm. MATERIALS AND METHODS The study was approved by a National Ethic Committee. Overall, we acquired data from thirteen normal volunteers. All subjects were scanned with echo-planar functional magnetic resonance imaging on a 3 Tesla whole-body MRI scanner (Bruker, Germany). In order to investigate the effect of the acquisition mode on the detection of activation and on the frequency content of the BOLD signal, a visual experiment was used. The experiment was performed with two different values for TR (3 and 6 seconds), corresponding to the acquisition of the whole brain and half of the brain volume respectively (see Acquisition parameters section). Simulation of the effect of missampling the haemodynamic response due to movement: The simulation study was performed as follow. A synthetic haemodynamic signal was simulated in a cluster of contiguous voxels located in adjacent slices with TR = 3sec and TR = 6sec, whilst a movement of half a voxel in the direction perpendicular to the slice plane occurred at half the simulated acquisition time. The haemodynamic signal consisted of an activation signal with a cosine form, plus white gaussian noise with a signal to noise ratio of 5. Neighbouring slices were set such that they contained the same activation signal but sampled with a TR/2 time difference, for the interleaved acquisition scheme, and with a TR/20 time difference for the sequential acquisition scheme. The magnitude of the signal reconstructed after a simulated perfect realignment was compared to the original magnitude. The results show that the HR signal is decreased by 20% (TR= 3sec) and by 50% (TR=6sec) after spatial re-sampling due to realignment in the case of the interleaved acquisition simulation. For the sequential acquisition simulation, the HR signal

3

is only decreased by 5%, whatever the TR value. Stimulation paradigm: We used a visual stimulation paradigm designed to highlight cortical areas sensitive to motion. It consisted of 300 dots, 1.1 degrees in diameter, randomly scattered in a circular aperture 16 degree diameter, that were static or randomly moving. Dots were presented moving during 500 msec (Stimulus) and kept static during 14.5 sec (Control). This interstimulus value was long enough to get the maximum amplitude of response to stimulation (2). Furthermore, the paradigm frequency (1/15 Hz) was less than half the sampling frequency (1/6 or 1/3Hz) so that the fundamental frequency, would not suffer from aliasing. Although the actual TR was 6 sec, it was possible to achieve a pseudo-sampling rate of 3 sec because the paradigm cycle was exactly 2.5 TR (5, 9). It should be noted that this stimulus, although sensory, can be considered as rather weak and it was also a subject of interest to see how the detection of the BOLD response to this very short stimulus would compare to the more commonly performed block designed experiment. Acquisition parameters: For seven subjects, we acquired the whole brain volume with twenty six slices per scan (TR = 6 sec). For six subjects, only 13 slices were acquired per scan so that TR was 3 sec. For each subject, we collected three runs of 103 (TR = 6 sec) or 153 (TR = 3 sec) scans using a gradient-echo echo-planar sequence. Each acquisition run was performed in a different mode that was interleaved (INT), sequential (SEQ) or sequential with a 1mm gap (GAP). To keep as close as possible to the standard acquisition parameters used in brain functional imaging experiments, the radio-frequency pulses for slice selection were kept to their usual 3-lobe-sinc value. In the sequential modes, slices were acquired from bottom to top, following the arterial direction. We considered that, given the large TR we used, there was no significant contribution from in-flow effects in our data (10) (8). Slice thickness was 5 mm except in the sequential acquisition with a gap (1 mm) where thickness was set to 4 mm in order to keep constant an inter-slice distance of 5mm. The slice resolution was 64x64 pixels and the in-plane pixel size was 3.75 x 3.75 mm2. The visual stimulation was triggered at the beginning of the acquisition. The order of the acquisition mode was randomised across subjects. Data analysis: Data from each subject were analysed individually using SPM (11) (12). Data were realigned using a sinc interpolation. The haemodynamic response was modelled using Fourier components, as described in (5). As sine and cosine functions model any constant phase shift value between stimuli and haemodynamic signal onsets, detection did not depend on the slice position and therefore data analysis was equally sensitive to the three acquisition modes. For TR = 6 sec, the model only included a sine and a cosine function of period 15 sec as covariates of interest. For TR = 3 sec, the model also included the first order harmonic frequency. Thus, both models, denoted "Full model", included the highest frequency below the Nyquist frequency (see ). Long term trends were modelled by low frequency cosine waves and entered as confound effects (using the so called high pass filter with

4

cut-off frequency = 1/30 Hz). Global intensity effects were also covaried out. At TR = 3 sec, it has been shown that the residual noise is auto-correlated (for instance, see (13); (14)). This time dependency can be taken into account in the calculation of statistical significance by filtering the data (and the model) with a kernel large enough so that the original correlation would become negligible (12). However, filtering was incompatible with our study, since it would have masked the possible increased smoothing expected from the interleaved acquisition. Hence, we did not filter our time series for detection purpose. Statistical results are therefore biased if considered on their own, but comparisons between acquisitions remain valid. For completeness, we checked the effect of this filter on frequency content of the signal. The statistical significance of the haemodynamic response to stimulation was tested with F-maps. F-values represent, for each voxel, the amount of signal variability that is explained by the inclusion of functions of interest in the model (in our case, these are sine and cosine functions at the paradigm frequency), normalised by the residual variance computed when all effects (interest and no-interest) have been removed (see (15)). A threshold (uncorrected for multiple comparisons) of p = 0.001 was applied to the calculated F-maps. This p-value corresponds to F = 7.49 with degrees of freedom (df) equal to (2, 86) for TR = 6 sec, and F = 4.96, with df = (4, 114) for TR = 3 sec. In this study, we do not compare directly the datasets acquired with different TRs as too many parameters may account for the differences that could be found in the results. F-maps comparisons: For each subject and session, the number of voxels above these thresholds was compared between acquisition modes, as well as their mean and maximum F-values. The maximum F-value represents a regional score while the mean and number of suprathreshold voxels can be considered as global scores. These scores are complementary and together they should reflect the overall detectability of the response. Based on these scores, we performed paired t-tests to compare the three acquisition modes. We also tested the acquisition mode and the subject factor with a multivariate analysis of variance. It should be noted that these two analyses assume an underlying normal distribution. With a limited number of subjects (6 and 7), there is no dependable method to check that these assumptions are valid. We therefore rely on the robustness of these procedures and consider the results both quantitatively and qualitatively. Frequency content: To test for the effect of the slice acquisition order on the frequency content of the responses, we used the two following models in the Sequential and Interleaved data sets for TR = 3 sec. In the first model, denoted "FF model", we tested the fundamental frequency as a covariate of interest while the first harmonic was covaried out as well as the low frequencies. In the second, denoted "1stH model", we tested the first harmonic while covarying out the fundamental frequency and trends (see Table 1). With these two models, we could estimate, in each voxel, the amount of variance explained by either the harmonic or the fundamental frequency, for both the sequential and interleaved acquisition modes. These tests were performed on 5 out of 6 subject data sets for which the F-value computed with the Full model reached a p-value corrected for multiple comparisons of 0.1. The F-maps obtained from the two models (FF and 1stH models) were compared, across acquisitions, at the p = 0.001 threshold (F = 7.34, df = 2 ; 114).

5

RESULTS The estimated movement was found to be of similar magnitude in the 7 subjects at TR = 6 sec and in the 6 subjects at TR = 3 sec. No systematic bias in favour of a specific acquisition mode was observed with a maximum translation and rotation of 2 mm and 2 degrees respectively across subjects and sessions. Activation results: TR = 3 sec Activation foci related to visual motion stimulation were located in the expected regions of the cortex, namely the occipital pole (V1/V2) and the occipito-temporal junction (V5). Five subjects out of six showed substantial response, i.e. maximum F-values above F(4,114)=10, irrespective of the acquisition modality. The sixth subject presented only one small significant focus of activation in the sequential with gap acquisition mode. Although the absence of BOLD signal is often associated with subject motion, the movement estimated for this subject was not larger than that estimated for the other subjects. Table 2 summarises the mean and standard deviation across subjects of the three scores for the interleaved (INT), sequential (SEQ) and sequential with gap (GAP) acquisitions. Testing for the acquisition mode factor in a multivariate analysis of variance (MANOVA) suggested a possible effect of this factor with p = 0.163; and results of the paired t-test between every mode showed a tendency in favour of GAP versus INT and in favour of SEQ versus GAP, for the maximum F value score (see Table 3). The subject effect was found to be very significant with p < 10-6, revealing a much higher inter-subject variability than intra-subject (inter acquisition mode) variability. Inclusion of the rank factor in the MANOVA model did not change its results, showing that the order in which acquisitions were performed was not a significant factor. Activation results: TR = 6 sec Table 4 presents the mean and standard deviation of the three parameters (mean and maximal F-value, and number of voxels above threshold) and shows the p-values given by paired t-test when comparing the INT, SEQ, GAP acquisition respectively. Table 5 shows significantly better results for the INT acquisition over SEQ and GAP, for both the mean F-value (p