We have developed a dual-wavelength optical imaging technique to (in effect) simultaneously image cortical blood volume and oxygenation in alert behaving macaques. This technique involves switching rapidly between two wavelengths: 530 nm (green, equally absorbed in oxygenated and deoxygenated hæmoglobin, thus measuring total hæmoglobin concentration, ‘HbT,’ or ‘blood volume’) and 605 nm (red, absorbed about 5-fold more strongly in deoxygenated than oxygenated hæmoglobin, thus measuring ‘oxygenation’.10
). While imaging V1 in animals performing periodic visual tasks we observed a hitherto unknown stimulus-independent hæmodynamic signal that appeared to entrain to trial timing ().
Periodic fixation tasks evoke stimulus-independent, trial-linked signals even in the dark
To study this trial-related signal in isolation, we developed a task that minimized visual input while preserving trial timing. In an otherwise completely dark room, the animal was required to fix its gaze periodically on a tiny fixation point for juice reward (point size ~1-2 arc min, i.e. ~ 1-2 cone diameters). The fixation point stayed on continuously, switching between two equiluminant colours to cue the animal to ‘fixate’ or ‘relax.’ It was akin to seeing nothing besides one single twinkling star in an otherwise black night sky. Our two rhesus macaque monkeys (‘V’ and ‘S’) learned the task correctly as evidenced by their fixation patterns (). Both monkeys performed long sequences of correct trials, consistently holding fixation during ‘fixate’ periods and taking fixation breaks, if any, only during ‘relax’ periods.
On imaging V1 while the animals performed this task we observed robust hæmodynamic signals at the trial frequency even though the animals were in virtually total darkness and foveal V1, the only region receiving visual input from the fixation point, lay outside our imaging area. These periodic fluctuations were seen in both the ‘blood volume’ (530 nm) and ‘oxygenation’ signals (605 nm; ). They were accompanied by periodic changes in heart rate (HR)11
and systematic pupil dilation12
on trial onset suggesting a rhythmic state of alertness synchronized to each trial ().
We wanted to determine the relation between these trial-linked hæmodynamic signals and V1 neuronal activity. A crucial assumption in most brain imaging studies is that hæmodynamic signals are caused by local neuronal responses through a uniform underlying mechanism1-6
). In particular, brain images are routinely used to infer changes in local neuronal activity by fitting the imaging signal with some standard causal hæmodynamic kernel. To reveal neuronal mechanisms underlying V1 hæmodynamics we obtained both trial-related and visually evoked optical imaging signals concurrently with electrode recordings across V1 (Fig S1
, Table 1 in supplementary material
). At each site, in alternating blocks (20-40 trials each) while the animal performed the same fixation task, we either presented vigorous visual stimuli or no stimuli at all. For each data set we then used an optimization routine to calculate the causal kernel that ‘best’ fitted hæmodynamics to concurrent neuronal signals (Fig S2
), and tested whether this ‘best’ kernel could reliably predict hæmodynamics.
To get measures of neuronal activity for this analysis we separated the electrode recordings into multi-unit spiking (MUA) and local field potential (LFP. Fig S1
). As expected, visual stimulation evoked vigorous responses in both MUA and LFP (). The stimulus-evoked LFP responses could be empirically separated into two distinct frequency bands (). The high-frequency band (‘hi-LFP’: 66-130 Hz, avoiding 60 Hz), like MUA, showed crisp visually evoked responses. The low-frequency band (‘lo-LFP’, 10-56 Hz), also showed robust signal but with no apparent correlation with visual stimulation. Our empirically defined LFP bands match categories defined through prior work. The ‘hi-LFP’ matches a frequency band (‘high gamma’) shown to correlate well with stimulus-evoked spiking and hæmodynamics2,15,16
. The ‘lo-LFP’ – often separated into finer frequency bands15,16
– is believed to have a very different relationship with other brain signals15,16
. We therefore decided to test the three neuronal signal types independently, MUA, hi-LFP and lo-LFP, for their ability to reliably predict concurrently recorded hæmodynamics. These tests were conducted separately for ‘blood volume’ and ‘oxygenation’ signals.
Local neuronal activity predicts visually driven, but not trial-related hæmodynamics
Visually driven MUA and hi-LFP predicted the simultaneously recorded hæmodynamic signals very well both in amplitude and time course (, S3b-e
). Further, the optimal kernels obtained by fitting these signals were consistent in shape across all recording sites (, S3c, top
); kernels from any given experiment predicted visually evoked responses in all other experiments with almost equal accuracy, attesting to their remarkable reliability (Fig S4
). In sharp contrast, the same kernels, when convolved with dark-room MUA or hi-LFP, were uniformly poor at predicting trial-related hæmodynamics, in both amplitude and temporal correlation (, S3b-e
). The latter finding (R2
~ 0.08, MUA; 0.06, hi-LFP) specifically implies that there is no consistent temporal relation between predicted and measured hæmodynamics, independent of amplitude. This poor predictability was particularly striking since the trial-related hæmodynamic signal amplitudes were almost comparable to those of responses to vigorous visual stimulation (37% at ‘blood volume,’ 530 nm; 57% at ‘oxygenation,’ 605 nm. Fig S5
). To check whether trial-related hæmodynamics could still be predicted reliably by concurrent neuronal recordings but through kernels of a different shape, we fitted dark-room MUA and hi-LFP to dark-room hæmodynamics. These ‘best’ dark-room kernels were highly variable amongst recording sites and, again, consistently failed to predict trial-related hæmodynamics (, S3c-e, bottom
). The same overall pattern of results was seen for both ‘blood volume’ and ‘oxygenation’ signals (Fig S6
These results provide compelling evidence that visually evoked hæmodynamic signals are very well predicted by established measures of local neuronal activity (MUA, hi-LFP) through a causal kernel that is uniform across experiments. Such a model fails profoundly, however, to predict the trial-related signals. Therefore any neuronal mechanisms underlying trial-related hæmodynamics appear to be distinct from those typically assumed to underlie neurovascular coupling.
Unlike MUA or hi-LFP, lo-LFP – whether treated as a whole or separated into finer frequency bands – failed to show any consistent relationship with hæmodynamics. These signals gave highly variable ‘optimal kernels’ when fitted with concurrent hæmodynamics either under visual driving or in the dark, with uniformly poor predictions of hæmodynamics (Fig S7, S13
Next, we characterized the novel trial-related hæmodynamic signal in terms of its temporal relation to trial timing. To determine whether our observed signals are linked specifically with trial timing and not a result of some unrelated intrinsic oscillatory process17
, we examined how the signals adapted to different trial periods. Our results provided compelling evidence that the signals are linked predictively to trial onsets. This was seen both in the signal shapes at each trial period and their anticipatory timing on switching trial period.
We found that the trial-related signals stretched elastically to match each tested trial period (, S8a
. Tested 6-sec to 30-sec trial periods). In particular, the shape of the ‘blood volume’ signal always stretched so as to start darkening (increasing hæmoglobin) during the ‘relax’ period, – before the onset of the next trial – reaching a peak darkening close to the onset of the next ‘fixate’ period (, ). This elastic pattern of trial-locked hæmodynamics – in which signals begin changing prior to trial onsets – cannot be explained by mechanisms that involve a causal kernel triggered on trial start. This can be demonstrated by comparison with responses to (brief, intense) visual stimulation of the same duration as the ‘fixate’ period, where the stereotyped response shape, with abrupt onset and fixed width following stimulus presentation, is independent of trial period (Fig S8b
; quantitative model, Fig S8c
). The trial-related signal is thus unlikely to be due to neuronal signals active only during the cued ‘fixate’ period (e.g. the presumed time course of ‘attention’18
Trial-related hæmodynamic signals entrain to anticipated trial onsets, stretching to conform to the trial period
On switching trial timing unexpectedly after the monkey had established a rhythm of 10-20 correct trials at a given period, hæmodynamic signals continued to oscillate at the earlier period for a couple of trials before entraining to the new one (). This occurred even though the animal himself picked up the new trial pace immediately, holding and breaking fixation at the new rhythm right after the switch (i.e. clearly having noticed the new pace of fixation cues). Thus, on switching from short to long trials the measured signals showed a peak darkening at the short trial spacing even though the animal was fixating correctly at the longer period (). Similarly, on switching from long to short trials the cortical signal continued at its prior slower pace for one long period, overriding the first few short trials (). The response shape observed on transition trials closely resembled pre-transition responses for the duration of the pre-transition trial period, while being very poorly matched to the post-transition trial shape suggesting that the underlying neuronal mechanism continued to ‘anticipate’ the pre-transition trial timing (, S9
). Further, the trial-related signal timing was correlated specifically with trial onsets and not with reward19
– the peak darkening position remained unaffected on delaying the reward associated with each trial (Fig S8d
Finally, images of the cortical surface suggest that the trial-related signals involve the local vasculature rather than being a systemic trial-locked autonomic (e.g. cardiac) response20
. These images revealed a dramatic contraction-dilation cycle in V1 arteries, evidenced by a prominent brightening, followed by darkening of the arterial walls relative to the ‘parenchyma’ baseline (, S10
. Fig S11
indicates how arteries, veins and ‘parenchyma’ are distinguished and how the artery signal is measured). This arterial signal had a timing that closely matched the overall timing of the mean ‘blood volume’ with peaks of arterial contraction and dilation coinciding with peaks of brightening (decreased hæmoglobin) and darkening (increased hæmoglobin), respectively (). The arterial cycle stretched elastically to fit trial periods, matching the shape of the mean signal (). Further, on switching trial periods the arterial cycle showed an anticipatory dilation well synchronized with the anticipatory increase in ‘blood volume’ seen in the mean signal (). This local arterial cycle may thus be the specific mechanism generating trial-related increases in V1 ‘blood volume’ in anticipation of visual tasks. Further, the arterial cycle is seen in V1 only for visual tasks and is likely not a passive consequence of trial-locked changes in heart rate or blood pressure20
. We found no V1 arterial pumping or trial-related changes in V1 ‘blood volume’ in a periodic auditory control task despite the presence of periodic changes in heart rate and pupil dilation very similar to those seen in our visual task (Fig S12
Mean ‘blood volume’ signal is closely matched, temporally, by V1 arterial contraction-dilation cycle
Our findings have two major implications, one for the interpretation of brain imaging21
, and the other advancing our knowledge of brain mechanisms underlying anticipation. First – the interpretation of fMRI22
, e.g. through general linear modelling23
, typically makes the crucial assumption of a uniform linear predictive relationship between neuronal and hæmodynamic signals. We show that this model is valid for visually evoked signals, but that it fails profoundly to predict another class of signals, of almost comparable magnitude and behaviourally linked structure. These results raise the further possibility that there may be other, hitherto uncovered exceptions13,14
to the assumption that hæmodynamic signals uniformly imply equivalent underlying neuronal activity. Second – the predictive timing and arterial contraction-dilation cycle that we observe in the trial-related hæmodynamic signal suggests that it could reflect a novel anticipatory brain mechanism. This mechanism could play the role of preparing cortex for anticipated tasks by bringing additional arterial blood in time for task onsets. The question of the mechanism driving this signal (e.g. distal neuromodulatory control of cerebral arteries?) as well as its functional consequences remains a challenge for future investigations.