Can current fMRI techniques reveal the micro-architecture of cortex?

To the editor

In a recent paper in Nature Neuroscience, Kim et al.¹ claimed to have visualized orientation columns in the visual cortex of cats by means of high-field, spatially resolved functional magnetic resonance imaging (fMRI). This would represent a striking technical advance indeed, if their mapping method were robust to noise and if the recorded patterns were reproducible within a given experimental session. A careful examination of their data, however, suggests that this may not be the case. Because the maps obtained in this study rely on a specific phase of the hemodynamic response, I will begin by briefly reviewing a few basic principles.

When a stimulus of a few seconds' duration is presented while T2^★-weighted images are being continuously recorded, the evolution of activation has at least two phases: an increase in intensity, usually referred to as overcompensation, and an undershoot with a long time constant following the offset of the stimulus². In addition, some fMRI time series show an initial, transient signal loss (dip) starting almost immediately after stimulus presentation and lasting for about four to five seconds^3,4,5. This peculiar bi- or triphasic evolution of the stimulus-induced activation occurs because the fMRI signal is sensitive to hemodynamic changes rather than to changes in neural activity per se. The commonly used blood oxygen level-dependent (BOLD) contrast signal, for instance, reflects changes in blood volume, flow and oxygen concentration. The early dip is thought to reflect a localized increase of deoxyhemoglobin in the parenchyma, whereas the later positive phase is believed to stem from an increase in cerebral blood volume and flow that extends considerably beyond the site of increased neural activity and tends to overcompensate for the increase in oxygen consumption⁶.

Kim et al.¹ generated maps of orientation columns in area 18 of the cat by exploiting this localized dip. Specifically, they presented cats with oriented, drifting gratings, and determined the stimulus-induced activation by monitoring any negative signal the absolute value of which was greater than (0.5+1s.ds)% at 2 seconds after stimulus onset (Fig. 1d of ref.1). They also obtained a conventional activation map (Fig. 1c in ref. 1) by correlating the time course of the hemodynamic response with that of the visual stimulus under exactly the same stimulus conditions in the same session.

Assuming that the dip is due to a local increase of deoxyhemoglobin, one would obviously expect every single time series exhibiting this early negativity to exhibit the subsequent overcompensation as well. In fact, optical imaging experiments have never detected an initial negativity that was not followed by the typical overcompensation (A. Grinvald, personal communication). Furthermore, if the dip reflects—as it should—a signal that is spatially restricted within the parenchyma, one would also expect it to occur almost exclusively outside the areas transpierced by major drainage vessels, such as the sagittal sinuses; this is particularly true of the very early (two-second) phase of the negativity that was analyzed by Kim et al.¹ Yet an exact superposition of Figs. 1c and 1d of ref. 1 shows that the above assumptions do not hold for their data (Fig. 1 below). The areas showing a dip largely overlap with the sinuses, and a dip signal can be also detected in areas showing no statistically significant conventional activation. The former is puzzling because studies in humans by other investigators from the same laboratory⁵, as well as studies in monkeys⁴, have shown that large vessels like the sinus show only positive signal changes. The latter is also puzzling, because the dip is far smaller than the conventional activation signal and should therefore be more difficult to detect in noisy images.

Figure 1: The activity derived by detecting the initial dip superimposed on the activity mapped by correlating the time series of each voxel with the stimulus time course (data from ref. 1).

White contours indicate some dip-determined areas that completely overlap with the sagittal or horizontal sinuses. Black contours indicate some other dip-determined activity in regions showing no significant positive activation.

A logical conclusion from these observations is that the initial dip, especially when analyzed with the simple statistics used by Kim et al.¹, yields maps that are extremely sensitive to noise and that may have very little to do with the actual columnar architecture of the cortex. An obvious test for reliability would be to repeat the acquisition of maps for a given orientation within a single session and to ensure that they are reproducible. Unfortunately, the authors have not reported any such tests. It should be noted, however, that repeated acquisitions in our own laboratory (unpublished—and ultimately unpublishable—data) using almost the same hardware, the same spatial resolution and very similar postprocessing techniques failed to reveal an acceptable consistency in dip-derived ocular dominance maps of the monkey visual cortex. Patterns that resembled columnar organization were found occasionally, but they were not reproducible within a single session and were therefore discarded.