High-field fMRI unveils orientation columns in humans

Reproducibility and Reliability of the Maps.

To test the reproducibility of the phase maps, several approaches were explored. Initially, within a session, the data were separated by odd and even scans and the respective phase maps were compared on a pixel-by-pixel basis [R = 0.62 and R = 0.66 for subjects 1 and 2, respectively (see Fig. S3 a and b)]. In an additional experiment, to further establish the repeatability of the observed orientation specific responses, we alternated the stimulus between a CW and a CCW sweep through the different orientations (see Fig. 4d) and checked reproducibility, accounting for the reversal of the stimulus presentation (R = 0.40) (Fig. S3c). This reproducibility is expected to be worse because of ambiguities introduced by the hemodynamic delay.

Grouping the voxels in a 90° bandwidth results in two complementary phase maps (Fig. 4a). Reproducibility of such binary maps within a session and between days is illustrated in Fig. 4 a–c. When such maps are obtained with the stimulus traversing CW vs. CCW (Fig. 4d), these binary maps are expected to largely, but not completely, reverse (Fig. 4 a vs. b). The reversal is not expected to be complete because of the hemodynamic delay. When the stimulus traverses CW, the hemodynamic delay results in a small shift in the observed orientation preference. Because the hemodynamic delay is ≈6 s and one cycle was 60 s, this shift is ≈10% of the orientations sampled (≈18°). When the stimulus rotates CCW, the shift in the depicted orientation preference is in the opposite direction (41). Despite this, the comparisons in Fig. 4 a and b show large areas of reproducibility.

The reversal of the stimulus traversal (CW vs. CCW) should also result in the reversal of pinwheel directions. In other words, pinwheels, where the orientation preferences are arranged either in the same or opposite rotational direction as the traversal of the stimulus, must reverse when the stimulus traversal direction is reversed. This is illustrated in Movie 3. If the pinwheel patterns are simply artifacts generated from noise, this organized reversal of direction cannot take place.

As a further test of the reliability of the observed spatial patterns, we used the Fourier magnitude analysis (see Methods and Fig. S2), which we previously used for spatial characterization of orientation preferences. However, instead of quantifying the spatial profiles of voxels with similar fMRI phases we used random spatial distributions of the orientation-specific phases. This analysis, unlike when the correct spatial information of the calculated phases is used, resulted in a lack of any spatial coherence in the Fourier magnitude as one traveled away from a given column (see Fig. S2). Finally, to demonstrate the relative strength of the response at the stimulus frequency, we plotted the fast Fourier transform (FFT) of the fMRI time course in the gray matter region of interest (ROI) and compared it with the FFT of temporally scrambled images from the same ROI. The Fourier spectrum distribution of the fMRI time course demonstrated a peak at the stimulus frequency. There are other peaks that can be ascribed to coherent fluctuations in the fMRI time series arising from low-frequency oscillations, cardiac cycle, respiration, and consequent harmonics or folding effects when multiple sources with different frequencies affect the time course. Coherent signal modulations, including the major peak appearing at the stimulus frequency, are eliminated in the temporally shuffled images, demonstrating that their appearance is not attributable to random noise (see Fig. S4).

fMRI sensitivity is paramount to resolving high-resolution functional patterns, as is done in this work. Here, each study comprised of eight to 10 scans (≈7 min per scan) with the subject in the scanner for 2–2.5 h. Generating the functional maps can be performed, however, with much less data averaging, depending on how still the subject remains throughout the scan. For example, in subject 2, we generated two such high-resolution maps (i.e., CW and CCW scans) from the data acquired during one session. Thus, temporal stability, and not fMRI sensitivity, is the limitation in these acquisitions.

Preferred Orientations.

A standing debate exists on whether there is a cortical preference (and/or global over-representation) to specific orientation(s). Previous psychophysical and behavioral studies in humans and animals have suggested or predicted either no spatial bias (7, 27, 42–46) or preferences toward both orthogonal orientations (0° and 90°). In the present data, when voxels were binned with a 30° bandwidth around 0°, 45°, 90°, and 135° orientations, a statistically significant (P_subject1 < 0.03, P_subject2 < 0.07; see Fig. 5) over-representation (i.e., more voxels preferring) of orientations around 90° (vertical orientation with the grid moving horizontally) was observed in both subjects. The error bars in Fig. 5a represent errors between scans within the same subject, and the asterisks represent the significance of the difference between the number of voxels at a given orientation vs. the number of voxels at 90°. To assess whether this over-representation of 90° represented a larger number or larger width of 90° columns, we investigated the spatial autocorrelation of the Fourier magnitude distribution of voxels with similar phases (see Fig. S2). This analysis revealed that column widths or spacing were not statistically different for the 4 phase bins; however, there was a higher regularity or density of columns representing vertical orientations in both subjects. This finding resembles the oblique effect (46–48), which has been demonstrated in the literature through imaging and psychophysical data. The oblique effect refers to greater behavioral sensitivity and larger BOLD changes to both vertical and horizontal orientations compared to oblique (45° and 135°). However, this effect was observed when spanning the entire visual field (all of V1) (46) as opposed to our restricted region. Such significant differences in the amplitude of the BOLD responses could also explain our bias, such that detection is affected in the phase analysis. Such amplitude differences, however, were not observed in our data.

Orientation Tuning Curves.

Under our experimental conditions, the functional contrast, i.e., signal change for active vs. nonactive columns, was found to be 67.5 ± 8% and 41 ± 5% for adjacent and complementary columns, respectively (Fig. 5b). The ratio of signal changes (nonactive/active) in voxels of a certain orientation preference when experiencing the orthogonal orientation (e.g., 0 vs. 90 or 45 vs. 135°) was ≈0.4, consistent with what was observed in the cat by using cerebral blood flow (CBF)-based fMRI (23). Although no statistical difference in tuning bandwidth was observed between columns, there was a small tendency for orthogonal orientations to better resolve oblique orientations (r = 0.4), than the reverse (r = 0.1). Interestingly, although there is a tendency of more cortical area being devoted to orientations around 90°, the average increase in fMRI contrast in an “active” column was relatively similar (BOLD change: 0.94 ± 0.22%, subject 1; 1.45 ± 0.27%, subject 2) across all orientations.

Imaging Resolution and the BOLD Point Spread Function.

We recently reported that when avoiding large vessels, the point spread function (PSF) (Gaussian, full-width half-maximum) of GE BOLD signals at 7 T has an upper limit of 2 mm (49), consistent with a GE-BOLD PSF of 1.6 mm measured in the cat^§. When the same approach was used, measurements in the cat yielded a PSF of ≈0.7 mm for both SE-BOLD and CBF-based fMRI^†, in agreement with the ability of CBF-based fMRI to yield single-condition orientation maps in the cat (23). Thus, the PSF of the SE-BOLD response should be sufficient to image orientation preference. However, even with GE-BOLD fMRI, it would be feasible to image regular and repeated structures, such as ODCs and orientation preference, provided that large vein areas and discontinuities encountered at boundaries are avoided. This is frequently unappreciated and consequently leads to confusing arguments about the feasibility of “columnar” imaging by fMRI with the different techniques. In imaging of periodic structures, the PSF of the imaging modality affects mainly the CNR, i.e., the relative signal amplitude of the active column relative to its complementary column, and not the ability to accurately define the spatial structure of the different columns. Of course, a low CNR ultimately affects the accuracy of the maps because the boundaries or even the existence of the columns cannot be accurately detected because of the noise. However, if CNR is high enough, then even a method with a larger PSF than the structure to be imaged can yield a map of the structure provided that it is periodic and the PSF is spatially uniform over the area imaged. This concept is illustrated using simulations in Fig. 6 with a 1D structure (2 mm/cycle), resembling human ODCs. Unlike the case of imaging just one column in a single-condition paradigm (Fig. 6a), the differential image of the repeated and alternating complementary columnar structure cancels the inaccuracy introduced by the PSF so as to define the boundaries of the structures accurately, albeit with different contrast (Fig. 6 b and c). The same would be true for multiple, regularly organized representations with similar response properties as in orientation preference mapping. Another advantage of using a technique with a small PSF occurs when the PSF and/or magnitude of the responses is not spatially uniform but has local variance. In fMRI, this is the case because of the presence of large veins distributed sparsely and irregularly, which yield large false “functional” responses especially in GE-BOLD (2). In this case, whereas cancellations in differential maps will yield columnar maps in some regions, in others, this cancellation will not be perfect, leading to obscuration of the underlying structure. This has been observed in ODC mapping with GE-BOLD (21). In SE-BOLD at high fields or CBF-based fMRI, this problem is minimized or even eliminated because large vessel contributions are substantially reduced.

Nevertheless, structures with 1-mm dimensions can be mapped in areas void of large vessels, as demonstrated previously in humans (19, 21) and animals (50). In animal models, this is particularly feasible because the geometry of the brain permits imaging in a slice that does not contain surface vasculature (50). Therefore, our data are consistent with the increasing understanding of resolution limits of fMRI and the growing body of animal data that demonstrate the feasibility of columnar mapping with fMRI.