Vascular contributions to pattern analysis: comparing gradient and spin echo fMRI at 3T

Abstract

Multivariate pattern analysis is often assumed to rely on signals that directly reflect differences in the distribution of particular neural populations. The source of the signal used in these analyses remains unclear however, and an alternative model suggests that signal from larger draining veins may play a significant role. The current study was designed to investigate the vascular contribution to pattern analyses at 3T by comparing the results obtained from gradient and spin echo data. Classification analyses were carried out comparing line orientations in V1, tone frequencies in A1, and responses from different fingers in M1. In all cases, classification accuracy in the spin echo data was not significantly different from chance. In contrast, classification accuracies in the gradient echo data were significantly above chance, and significantly higher than the accuracies observed for the spin echo data. These results suggest that at the field strength and spatial resolution used for the majority of fMRI studies, a considerable proportion of the signal used by pattern analysis originates in the vasculature.

Fig. 1

Schematic representation of the task. In each block of trials, participants associated one of two oriented gratings with one of two sine wave tones. On each experimental trial, participants were presented with a single grating/tone pair and asked to indicate whether the pair was correct or incorrect according to the current mapping.

Fig. 2

Mean classification accuracy rates for represented and non-represented stimulus dimensions in each region of interest for each type of acquisition sequence. Represented stimulus dimensions (i.e., tones in A1, fingers in M1, and orientations in V1) were discriminated with above chance accuracy in the gradient echo data but not in the spin echo data. * = 1 sample t-test against chance (50%), p < 0.05. Error bars represent 1 standard error of the mean (calculated across participants).

Fig. 3

Mean classification accuracy rates in unsmoothed data and at three different levels of spatial smoothing. Smoothing was carried out using Gaussian kernels of either 4, 6 or 8 mm FWHM. The effect of smoothing varied across regions, producing a decrease in classification accuracy in A1 and M1, but an increase in accuracy in V1. Error bars represent 1 standard error of the mean (calculated across participants).

Fig. 4

Event related time courses for each of the four sequence types plotted alongside the haemodynamic response function (HRF) obtained using the default parameters in SPM5. In all cases there was a more rapid decline from peak response in the empirically derived response functions than in the canonical HRF. In addition, the spin echo data showed a shorter peak latency in V1 and M1. For clarity of display, the canonical HRF function has been arbitrarily scaled to have a peak response 10% greater than the maximum height of the empirically derived event related time courses.

Fig. 5

Mean classification accuracy results obtained using basis functions modelled on the empirically derived event related time courses shown in Fig. 4. In A1, this approach produced an increase in mean accuracy of 4.72% in the standard resolution GE data, 4.75% in the high resolution GE data, 2.67% in the standard resolution SE data, and 0.49% in the high resolution SE data. The corresponding figures in M1 were 2.61% (standard GE), 0.3% (high GE), and 1.52% (standard SE). In V1, there were increases in the GE data (standard resolution = 1.36%, high resolution = 2.62%), but slight decreases in the SE data (standard resolution = −0.34%, high resolution = −0.35%). * = 1 sample t-test against chance (50%), p < 0.05, ⁺p = 0.054. Error bars represent 1 standard error of the mean (calculated across participants).