The current report describes the effect of several factors on multivariate pattern classification in primary sensory and motor areas. One of the most salient results was the significant difference between classification accuracies obtained from data collected using a gradient echo sequence and those obtained from data collected using a spin echo sequence. While it was possible to detect differences between line orientation in V1, tone frequency in A1 and finger in M1 using the gradient echo sequence, classification accuracies were not significantly different from chance in the spin echo data. It is possible that improvements in the signal-to-noise ratio (for example, by using surface coils rather than a birdcage head coil), or contrast-to-noise ratio (for example, by including more observations) might allow above chance classification in the spin echo data, although as any increase in the signal- or contrast-to-noise ratios would also be beneficial for the gradient echo data, the difference in accuracy between the two sequence types should remain.
From a methodological perspective, the current results suggest that the gradient echo sequence may be preferable for multivariate pattern analyses, at least at the field strength and spatial resolution currently used for the majority of fMRI studies. The results also suggest that the signal used in pattern analyses contains a considerable contribution from larger draining veins, since classifier performance showed a marked decline when this signal was reduced in the spin echo data. In this respect the result is consistent with the findings reported by Shmuel et al. (2010)
examining the spatial distribution of information about ocular dominance in V1. By calculating the discriminative power of individual voxels and comparing the results with the probability that those voxels belonged to regions containing macroscopic blood vessels, the authors demonstrated that peaks of discriminative power were present in voxels containing larger draining veins.
While the extra-vascular signal around larger draining veins should have been eliminated in the spin echo data, previous work suggests that with parameters similar to those used in the current study, there would still be some signal originating from inside larger veins (Norris et al., 2002; Duong et al., 2003; Jochimsen et al., 2004
). This intravascular signal is expected to be weaker than the extra-vascular signal however, and the current results suggest that it was not strong enough on its own to provide a useful source of information. Further work using acquisition sequences that suppressed intravascular contributions (for example comparing gradient echo data with and without diffusion gradients) would be necessary to determine the relative contributions of intra- and extra-vascular signals, in particular whether either source in isolation could provide sufficient signal for reliable pattern classification. Similarly, comparisons of gradient and spin echo data in the presence of diffusion gradients (or at high magnetic field strengths) would be necessary to determine the effect of completely eliminating any signal from larger vessels.
While the current results demonstrate the importance of the vascular signal at 3T, they do not rule out the possibility of gray matter or microvascular contributions to pattern analysis. In the study described above, Shmuel et al. (2010)
found that discriminative power was also present in voxels that contained mainly gray matter, and it is possible that multiple sources of information exist. Importantly, Shmuel et al. (2010)
carried out their study at 7T, and it is likely that the relative contributions of gray matter and draining veins will depend on magnetic field strength (Duong et al., 2003
). The current results suggest that at 3T pattern information is provided mainly by signals from larger draining veins. At higher field strengths, however, it is possible that the relative increase in the strength of the microvascular signal (Lee et al., 1999; Duong et al., 2003; Jochimsen et al., 2004
) may produce a situation that more closely resembles the biased sampling model, with patterns of spatially localised signals originating from different populations of cells. In line with this idea, another study carried out by Shmuel et al. (2007b)
at 7T reported equally high classification accuracy for ocular dominance columns in both spin and gradient echo data.
In addition to the effect of sequence type, the study also investigated the role of spatial smoothing on classification accuracy, and the results present a complex picture, with the precise effects of smoothing depending on the region and stimulus dimension in question. The results from V1 are consistent with those recently reported by Op de Beeck (2010) and Gardner et al. (2006)
in demonstrating that spatial smoothing does not harm classifier performance in V1. Indeed, in the current study, classification of orientation in V1 actually appeared to be enhanced by spatial smoothing. This effect was specific to V1 however, and increasing the size of the smoothing kernel reduced the accuracy of tone classification in A1 and finger classification in M1.
On first inspection, the different effects of smoothing appear somewhat surprising given what is known about the functional organisation of the three regions. In human primary auditory cortex, functional imaging has revealed mirror symmetric tonotopic maps where particular frequency bands are represented by populations centred on locations separated by several mm (Formisano et al., 2003
). The topography of M1 appears to be organised on a similar spatial scale, with individual fingers represented by partially overlapping populations centred on foci that are approximately 2–4 mm apart (Indovina and Sanes, 2001; Dechent and Frahm, 2003
). In contrast, representation of orientation in V1 appears to be organised into much smaller functional units (~ 750 µm in width) that are widely distributed over an area of several square cm (Yacoub et al., 2008
Assuming that pattern analysis measures differences in the distribution of these functional units, the matched filter theorem would predict that V1 should be more vulnerable to smoothing than that in either A1 or M1, contrary to the observed pattern of results. Crucially, representation in V1 is periodic, compared to the relatively localised representations in A1 and M1 where individual tones or fingers are maximally represented at distinct spatial locations. The repeating pattern in V1 raises the possibility of either a pre-existing low frequency component to the information (Op de Beeck, 2010
), or that high frequency information was aliased into low frequency components (Kriegeskorte et al., 2010
). In either case, smoothing could have produced a relative amplification of low frequency information in V1, while diluting the more localised information present in A1 and M1. Interestingly, there was some evidence that both A1 and M1 showed an initial increase in accuracy when a small amount of smoothing (4 mm FWHM Gaussian kernel) was introduced. Given the predictions of the matched filter theorem, it would be interesting to test whether the optimum level of smoothing in these regions depends on the particular tones or fingers that are compared, in particular whether comparisons between tones or fingers that are represented by populations with greater spatial separation benefit from more smoothing.
Another possibility for the pattern of results in V1 is that there was a greater representation of one particular orientation, producing global differences in signal between the two conditions. Previous results suggest that there may be differences in the representation of horizontal and vertical orientations, although the picture is inconsistent, with some studies reporting an over-representation of vertical orientations (Yacoub et al., 2008
), some an over-representation of horizontal orientations (Serences et al., 2009
), and others reporting similar levels of signal change for horizontal and vertical orientations (Furmanski and Engel, 2000
). In addition, the univariate analyses carried out in the current study showed that there were no global differences in activation between the two orientations.
The current study focuses on primary sensory and motor areas, and the relative contribution of micro- and macro-vascular signals in regions of the brain with less differentiated neural topographies remains an open question. Together with several other recent reports however, the current findings highlight the importance of taking both into account as potential sources of information. One relatively low cost strategy for future studies focussing on areas where the underlying neural topography, or patterns of neuro-vascular coupling, are unclear could be to use a dual echo sequence with near simultaneous acquisition of both gradient and spin echo data in each TR (Bandettini et al., 1993
). At high field strengths this could provide a direct estimate of the vascular contribution to pattern separation. This type of approach could also be used to give an estimate of the mean vessel size within each voxel (e.g., Jochimsen and Moller, 2008
), and a comparison of classification results from voxels containing different vessel sizes could provide further important insights into the source of pattern separation.