The chronoarchitecture of the cerebral cortex

1. Chronoarchitecture: the principle of functional independence and natural conditions

One hundred years after their derivation, the cytoarchitectonic and myeloarchitectonic maps of the cerebral cortex derived by Brodmann (1905, 1909), Campbell (1905), Vogt & Vogt (1919), von Economo & Koskinas (1925) and others are, remarkably, still in use. This may seem surprising given that there are many subdivisions that they have not revealed and given the advent of new imaging techniques of infinitely greater sophistication. Perhaps one important clue to their success lies in the fact that they do not hypothesize about functions. Instead, their only hypothesis is that anatomical subdivisions will reflect functional subdivisions. By contrast, activations obtained in neuroimaging experiments depend heavily on the exact nature of the stimuli used, and their interpretation is heavily hypothesis-driven (Friston et al. 1995; Frackowiak et al. 1997). Moreover, the same cortical region (e.g. parietal cortex) may be activated in a variety of different ways, including more or less of its functional subdivisions, depending upon the stimulation paradigm used. While such a picture may well reveal the functional complexity of the parietal cortex, it also makes it difficult to subdivide it in a compelling way. The challenge of mapping cortical subdivisions thus lies in finding the right anatomical or functional markers. Unfortunately, there are relatively few functional maps provided by modern neuroimaging experiments in the human that can be reproduced reliably, such as that of area V5 or of retinotopically mapped visual regions, while extent and functional interpretation of activations in large parts of the cortex, such as parietal, temporal or frontal regions, vary with the stimuli and with changing insight. Equally, despite their reliability, cytoarchitectonic maps (Brodmann 1909) or other anatomical mapping techniques like recent attempts of metabolic or molecular maps (Livingstone & Hubel 1984; Zilles & Clarke 1997) have provided relatively selective cerebral maps that do not nearly reveal all the functional and structural subdivisions of the cortex or that apply only to subregions. This is perhaps best illustrated in the case of cytoarchitectonic area 18, which has been shown in subsequent physiological and anatomical studies to contain at least four further areas that differ significantly in their functions. Among these are areas V2, V3, V3A and V4 in the monkey (Zeki 1978). Modern multi modal databases that combine maps derived from several methods are more powerful as they combine the strengths of individual methods, and together with large databases and new approaches, such as computer-aided high-resolution anatomical mapping, will without doubt provide more powerful tools in the future (Mazziotta et al. 2001; Roland et al. 2001; Van Essen 2002).

2. Anatomical maps and their relation to function

The concept underlying any anatomical, structural or architectural method is simple and enduring. First formally enunciated by Oskar & Cecile Vogt (Vogt & Vogt 1919), it merely states that architectural differences are indicative of functional differences and, conversely, that functional differences demand differences in architecture—the dominant and most successful theme in cortical studies. The enduring success of this approach is reflected in the continued use of the resulting brain maps today. Indeed, the approach has seen something of a spectacular comeback following the introduction of imaging studies and attests to the approach's conceptual soundness. Speaking in modern terms, brain function is implemented in the form of neuronal hardware. Thus, by definition, differences in function require differences in hardware, visible in terms of cell types, connectivity, synaptic and molecular structures. The principal methods used by the early cartographers of the cerebral cortex were those of cytoarchitectonics and myeloarchitectonics, the former revealing the manner in which cells are stacked upon one another in layers and the latter the pattern of myelination in different zones of the cerebral cortex (Flechsig 1901; Brodmann 1909). Irrespective of the shortcomings of these methods, they had two over-riding advantages. First, whatever architectural boundaries they delineated could be assumed to delineate functional differences. Second, the very nature of the anatomical approach, which is ‘blind’ with respect to the function, ensured a certain objectivity and longevity of the findings. This is especially apparent when compared with current, hypothesis-led imaging experiments, the results of which depend heavily on the particular paradigms used and rarely lead to an unchallenged attribution of function to a given area. With this supposition at its basis, cartographers such as Brodmann (1905), Campbell (1905), Vogt & Vogt (1919) and von Economo & Koskinas (1925) did not have to overly concern themselves with attaching a function to every subdivision that they managed to delineate. Instead, they hoped that future functional and clinical studies would validate their subdivisions. This has turned out to be true to some extent. Apart from the well-worn examples of the striate cortex and the motor cortex, there are several other architectural subdivisions that have proven to reflect, perhaps to an unexpected degree, a functional differentiation. One example is provided by the subdivisions of the parietal cortex. Although authorities may still disagree on whether the subdivisions of this cortex provided by Brodmann accurately reflect the functional picture that modern physiological evidence is providing, its separation from the rest of the prestriate cortex is reflected in its functional uniqueness. Equally, Brodmann's delineation of area 19 from area 18, at least in the upper reaches of the superior temporal sulcus, fairly accurately reflects the transition into area V5 or, to put it more accurately, into the V5 complex of areas.

Of the two techniques used by the early cartographers, the more successful and interesting has been the myeloarchitectonic method. This reveals the pattern of myelination in different areas. Among its successes are the delineation of an area in what used to be considered as ‘associational’ cortex that is myelinated at birth and coincides almost exactly with area V5 (Flechsig 1901; Watson et al. 1993), the sharp delineation of a boundary between areas 18 and 19, representing the transition to the cortex of area V5 as one moves rostrally across the cortex, and the delineation of several subdivisions in the fusiform gyrus (Lungwitz 1937), an early indication of a now well-established fact that the fusiform gyrus consists of several distinct areas that are apparently engaged in different tasks (Bartels & Zeki 2000a; Malach et al. 2002; Spiridon & Kanwisher 2002). However, there is another feature of myeloarchitectonic maps that has not been commented on to date. That is, that myeloarchitectonic maps, in theory, may give an indirect indication of the speed at which signals travel within or to a given region, based on the basic physiological knowledge that myelin enhances the speed of action potential propagation within the axon. Recent evidence confirmed this directly, suggesting that differential regional myelination yields constant latency irrespective of axonal distance, in this case between thalamus and cortex (Salami et al. 2003). The evolutionary reason for distinct myelination in distinct cortical regions is unclear—whether it may speed up signals of particular behavioural importance, or whether it is an attempt to synchronize percepts arising from regions with distinct processing delays is not known (see the hypothesis of perceptual synchronization; Bartels & Zeki 2004b). Whichever interpretation is true, myeloarchitectonic maps could be regarded, in a sense, as the anatomical reflection of chronoarchitectonic maps for the millisecond range and are probably related in part to the distinct signal arrival times observed across distinct regions, as well as the distinct, psychophysically determined, perceptual delays of different features (Moutoussis & Zeki 1997; Zeki & Bartels 1998a; Schmolesky et al. 1998).

However, modern studies have also shown the limitations of this anatomical approach. Perhaps one of the best-known examples is to be found in area 18 of Brodmann. While the cytoarchitectonic and myeloarchitectonic methods have not been able to show the subdivisions within area 18 that we now recognize, the fact remains that other techniques, for example, the technique of metabolic architecture (Wong-Riley 1979), have revealed a distinctive architecture for area V2—consisting of the well-known cycle of metabolically distinctive stripes (Livingstone & Hubel 1984)—and for area V3, consisting of a band of heavy metabolic activity (Jen & Zeki 1984). Moreover, functional studies have succeeded in associating distinct architectural subdivisions within individual areas such as V1 and V2—revealed by the technique of cytochrome oxidase architecture—with distinct functional groupings of cells (Livingstone & Hubel 1984; DeYoe & Van Essen 1985; Shipp & Zeki 1985; Xiao et al. 2003).

Thus, the statement that functional differences are reflected in architectural differences remains true today. We propose that the structural and functional borders can be inferred by the temporal marker of ATCs observed during natural conditions.

3. ICA applied to fMRI

4. Experimental comparison of ICA and SPM in controlled datasets

Before we used ICA to segregate functional subdivisions based on their ‘temporal fingerprints’, which we hypothesized to be optimally present during natural conditions, we had to convince ourselves that ICA was capable of doing so. In this section, we review the application of ICA to four fMRI datasets from controlled experiments which we also analysed using conventional SPM (Friston et al. 1995). This allows a direct comparison of the ICA results with those of SPM. These experiments include studies on colour vision, romantic love, spatial attention and object recognition (Zeki & Bartels 1999; Bartels & Zeki 2000a,b, 2004a). Figures 1–4 provide a comparison of the results obtained with ICA with those obtained using SPM. The equivalence of the results is evident. Figure 1 shows activity evoked by coloured abstract ‘Mondrian’ stimuli in the fusiform gyrus in the ventral cortex, which contains two separate colour-selective regions, V4 and V4α. The SPM analysis revealed adjacent representations of upper and lower visual fields in V4, which was not evident in V4α, while ICA simply isolated both colour-selective regions together in one IC (Bartels & Zeki 2000a). Figure 2 shows two of the regions that were specifically active when volunteers viewed pictures of their partner/loved one (in comparison to gender-, age- and familiarity-matched friends), namely a subdivision of the anterior cingulate cortex and the medial insula (Bartels & Zeki 2000b). The same regions were also activated when mothers viewed their own children as opposed to other acquainted children or their best friend (Bartels & Zeki 2004c). ICA isolated these regions in two separate components, indicating a differential involvement and the associated time courses reveal their specific activation in the attachment condition (Bartels & Zeki 2000b). Figure 3 shows activation related to spatial attention to either colour or motion in each of four visual quadrants. While the stimuli remained identical, spatial attention modulated activity in visual areas in a retinotopic fashion. SPM and ICA revealed identical results (Zeki & Bartels 1999). It should be noted that after one region has been isolated in one IC, the same region is unlikely to appear in other ICs again, mainly because of the ICA's constraint of spatial independence. Thus, only the SPM analysis of this dataset revealed that the colour-selective regions V4 and V4α, and the motion-selective region V5 were modulated not only by the spatial aspect of attention, but also by the feature attended to within the attended quadrant (motion or colour), which was not apparent in ICA (Bartels & Zeki 2000a). Thus, while ICA proved to deliver reliable results, classical statistical analyses allowed for additional statistical comparisons. Finally, figure 4 shows results from an object recognition study in which objects and their scrambled counterparts were either moving or stationary and were alternated with blank stimuli. Both ICA and SPM revealed indistinguishable results and the results obtained using ICA were highly reproducible across subjects (Bartels & Zeki 2004a). We found that the reason for the variability in the spatial extent of early visual areas in the ICs across the five subjects lay in oversplitting of this region into several ICs in some of the subjects. This could be easily detected using CMs that revealed early visual cortex as a whole, and would presumably be less likely to happen with recent optimized versions of ICA. More importantly, early visual areas (V1/V2) had consistently different activity time courses compared with those of the motion-selective area V5, even within a single condition: V1 showed adaptation while V5 did not, which was consistent across subjects. This provides a hint that apart from their distinct feature selectivity, distinct regions exhibit distinct dynamics in their response, which will work to our advantage in the ‘blind’ dissection of the cortex during natural conditions. We thus conclude that spatial ICA applied to fMRI data is a powerful and reliable way of segregating temporally related voxels. In fact, it is so successful that its results are qualitatively equivalent to those obtained using statistical mapping and may at times supersede them, as they demonstrate specificity of activity in separate areas or temporal correlations in networks of areas rather than correlation with experimental tasks. For the analysis of traditional, controlled datasets, however, SPM offers the advantage of explicit statistical testing of hypotheses, making it preferable over data-driven approaches. Indeed, several others share this view (Friston 1998; Calhoun et al. 2001a; Nybakken et al. 2002; Quigley et al. 2002). It is therefore questionable whether data-driven approaches such as ICA will remain widely used unless they are improved to enable more sound statistical inference (McKeown 2000; Calhoun et al. 2001b). Another difficulty lies in the selection of ICs because various non-neuronal signal fluctuations such as movement, breathing and heart-rate changes may potentially lead to ICs that are difficult to discern from ones relating to neuronally induced signal changes. In our view, hypothesis-driven analyses, such as SPM, are thus currently preferable for most imaging experiments. We see the key application of data-driven approaches, such as ICA, in the determination of spatial activity patterns based on their differential activity time courses in completely uncontrolled experiments. Irrespective of the precise experimental settings, we refer to them as chronoarchitectonic mapping, as they rely entirely on temporal markers.

5. The chronoarchitecture of the cerebral cortex

The perceptual richness and complexity of natural conditions that the brain evolved to deal with is likely to reveal the functional specialization of cortical regions optimally, as we explained above with the principle of functional independence. We approximated natural conditions by showing eight volunteers part of the James Bond film Tomorrow Never Dies (Bartels & Zeki 2004a). Note that this approach is diametrically opposed to that used in traditional fMRI experiments, which aim to differentially activate as few regions as possible by using as well-controlled stimuli as possible. Thus, the ‘blind’ nature of our approach required the development of some new techniques to obtain physiologically meaningful results, which we described in detail in our original publication (Bartels & Zeki 2004a). While the application of ICA to the free viewing fMRI data per se was identical to that in the traditional datasets, this time we could not use the onsets and timings of distinct conditions to select our ICs of interest. The selection of meaningful ICs is a major task in every ICA analysis, as there are as many ICs as original whole-brain scans, in our case up to 368 ICs per subject. Other groups have used PCA to reduce the amount of data before application of ICA, but we refrained from doing so because weak but physiologically important signals may get lost during this process. Firstly, variance of the fMRI signal due to neuronal processing can be smaller than that induced by artefacts such as breathing, head movement and so on. Low variance PCs may thus contain the data we are interested in keeping. Secondly, since some artefacts may be partially correlated with signal of neuronal origin, discarding PCs containing artefacts may also remove signal we are interested in due to PCA-imposed orthogonality. Thus, many of the ICs contain clusters of voxels that are not of primary interest to us in this context, such as ventricles, blood vessels and regions affected by head movements. Examples of such artefactual ICs are shown in figure 5. We used a conservative two-stage process to select ICs, thus ensuring that only ICs functionally involved in processing the stimulus were selected as determined by the significance of the intersubject correlations of their ATCs. In general, one can assume that the less stringent the selection criteria, the higher the probability of confusing some artefact with a functional subdivision. First, anatomical criteria such as location, extent and bilaterality were used to manually select candidate ICs, which reduced the original number to a few dozen per subject. In the second stage, we exploited anatomical and temporal correspondence of ICs between subjects. Functionally corresponding ICs of different subjects not only corresponded anatomically, but also showed significant inter subject correlations over time (see inter subject correlation and inter subject ranking in Bartels & Zeki 2004a). In addition, for each IC, a CM was calculated based on its most active voxel to confirm that the full spatial extent of the functional cluster was captured in the IC. This is a good method to check whether ICA super-fractionated a given area into small units (Bartels & Zeki 2004a, 2005).

6. Intra cortical correlations during natural conditions: a guide for anatomical connectivity?

An additional property apparent in some of the ICs was that they contained not a single cluster of voxels, but several. Voxels that are present in a single IC are temporally dependent in relation to their relative activity within that IC. Note that not all voxels which have high temporal dependencies necessarily need to be coisolated in the same IC given that spatial ICA applies its independence criterion only to the spatial domain, not to the temporal one. Thus, multiple clusters in ICs reveal functional connectivity (i.e. temporal correlations between distinct regions; Friston et al. 1993), but do not necessarily reveal the full network involved (Bartels & Zeki 2005). In order to reveal the full extent of the functional connectivity of a given region isolated in an IC, we took the BOLD signal of its most active voxel in each of the subjects to create CMs. CMs show the temporal correlations of all voxels in the brain with the seed voxel. These CMs either contained the same set of clusters as was already apparent in the IC or, in some cases, additional clusters. CMs were generally nearly as specific as ICs, containing a few clearly isolated clusters of voxels. This indicated firstly that much of the dependence ICs were based on was due to second-order correlations. Secondly, it showed that dependencies across the brain were highly specific (see figure 12). Lastly, we note that the correlation of the BOLD time course of the hottest voxel of an IC with that of the IC-ATC was, on average, 0.9 or higher. Most frequently, the secondary clusters in CMs were simply the homologues of a single region, such that both left and right homologues of a given region appeared in a single IC (see figure 6), which indicated a high correlation between them. We believe it is no coincidence that these particularly strong correlations were observed between regions that are known to have strong direct anatomical connections through the corpus callosum (Zeki 1970; Pandya et al. 1971; Clarke & Miklossy 1990; Bartels & Zeki 2005). This is in accord with the finding that the significant correlation between homologue regions during rest is diminished in patients with agenesis of the corpus callosum (Quigley et al. 2003). In addition, several ICs contained tertiary clusters, indicating correlations between distinct regions within the same hemisphere. We hypothesize that these, too, may be related to anatomical connections. We illustrate this on the example of the regions shown in figure 12a. It shows ICs containing a region overlapping with Wernicke's area. These ICs were isolated in every subject, and their associated ATCs had specific inter subject correlations indicating a common functional involvement. In most subjects, this IC contained an additional cluster in a region close to Broca's area. CMs revealed that Wernicke's area correlated in every subject with the same set of regions, including putative Broca's area (BA 45), another prefrontal/premotor region in the vicinity of BA 6, and BA 39 in the parietal cortex, at a threshold of r > 0.4 (see figure 12b). Parallel to the bilateral visual areas, these specific correlations within language-related regions reflect the known direct anatomical connections between them (Kaas & Hackett 2000; Petrides & Pandya 2002; Scott & Johnsrude 2003). Furthermore, as described in detail in Bartels & Zeki (2005), we observed that correlations between regions that are known to be anatomically connected increased during natural viewing, while correlations between unconnected regions decreased during natural viewing in comparison to rest. We illustrate this using the example of correlations between visual regions with language-related regions that de-correlated during natural viewing, while correlations within the language network increased (figure 14). The correlations were based on the BOLD signal extracted from the hottest voxel of ICs containing visual or auditory regions. This double dissociation of correlations during natural viewing and rest between connected and unconnected regions led us to propose that intra cortical correlations during natural conditions may be indicative of anatomical connectivity (Bartels & Zeki 2005). Especially noteworthy is the consistent involvement of BA 39 in CMs using Wernicke's area as a seed. While BA 39 was not classically considered to have direct connections with Broca's area and Wernicke's area, a recent study using DTI and track tracing methods revealed it to be a directly connected part of the language network (Catani et al. 2005). Our CMs revealed the same result, highlighting the huge potential of natural viewing CMs as hypothesis-generating tools for the mapping of anatomical connectivity. Apart from our experimental evidence, the relation of natural viewing correlations to anatomical connectivity does also make sense at a conceptual level. The brain evolved and develops during exposure to natural conditions. Its anatomical connectivity is therefore likely to reflect the degree of communication between distinct regions during these conditions. Thus, it seems plausible that intra cortical correlations reflect anatomical connectivity most accurately when they are measured during more natural conditions. This is not to say that correlations during these conditions invariably reveal anatomical ones. Firstly, one has to be aware of potential artefactual correlations (e.g. induced by head movements, breathing, changes in heart rate, etc.) combined with common vascularization. These artefacts affect all imaging studies and are particularly dangerous to those relying on temporal dependencies of any kind. Secondly, some anatomical connections may not be apparent at all through correlations, and some correlations may arise for other reasons. Methods relying on temporal correlations, of course, can only reveal anatomical connections that are active. Each area of the cerebral cortex has multiple outputs to different destinations, both cortical and sub cortical. It is not clear whether all these outputs are engaged when a given area is active. Indeed, previous studies have shown that distinct tasks or stimuli may lead to distinct degrees of functional connectivity (i.e. correlation) between regions (Buchel & Friston 1997; Friston & Buchel 2000; Friston et al. 2003). In view of this, we hypothesize that natural conditions are most suitable for eliciting correlations that reflect the underlying anatomical connectivity. Alternatively, one could say that the conditions should be as varied as possible in order to learn about all possible interactions of an area. Our results show that intra cortical correlations during natural conditions are more specific with respect to the known anatomical connectivity than those observed during the resting state, which have been implied to be indicative of anatomical connectivity previously (e.g. Lowe et al. 1998; Cordes et al. 2000; Hampson et al. 2002; Koch et al. 2002; Quigley et al. 2003; Young et al. 2003). As we are only beginning to understand the relation between the BOLD signal measured using fMRI and the underlying neural activity (Logothetis 2003), any inference on connectivity can only provide hypotheses rather than conclusive evidence. Despite this, such hypotheses are valuable, especially in the human brain, as we have only limited knowledge about its anatomical connectivity. DTI is a non invasive technique that is more directly based on anatomical connectivity by exploiting the diffusion preference of water molecules along myelinated fibre bundles (Parker et al. 2002). Unfortunately, while this method delivers impressive results, it is mostly limited to connections that originate in sub cortical nuclei or white matter. The numerous fibre crossings near the cortex are difficult to resolve within the dimensions of DTI-weighted image voxels, and thus make cortico-cortical tracing difficult, even though better solutions are in sight (Tuch et al. 2003) and first impressive results have been published as described above (Catani et al. 2005). We therefore envisage that DTI and connectivity patterns inferred on the basis of temporal correlations during natural conditions may be suited to complement each other for the charting of cortico-cortical anatomical connectivity in vivo. Our conceptual understanding and our experimental evidence suggest that correlations during natural conditions may be better suited as indicators of anatomical connections than other, more artificial, experimental settings (Bartels & Zeki 2005).

7. Conclusion

In this paper, we have reviewed a new approach to mapping functional subdivisions of the brain. It does not rely on anatomical markers nor on hypothesis-driven experiments. Rather, it relies on a marker that is present and unique to every functional subdivision of the brain; namely, its ACT during natural conditions (Bartels & Zeki 2004a,b, 2005). Chronoarchitectonic maps, therefore, mirror traditional ones. However, because they collapse all the known and unknown functional and structural properties of cerebral architecture onto a single and measurable factor, they may be easier to derive, even for uncharted regions. In addition, because the temporal fingerprint is a direct consequence of function, it also reflects functional similarity (or functional connectivity) and during natural conditions, this may be indicative of anatomical connections between areas (Bartels & Zeki 2005). The chronoarchitectonic approach therefore constitutes the logical development of one of the principal themes of cortical studies; namely, the attempt to identify distinct parts of the cerebral cortex and assign specific functions to each. It distinguishes and defines areas based on the timing of activity in them in relation to the timing of activity in other areas, without a priori knowledge about anatomical location or function. Its weakness, which it shares with other anatomical mapping techniques (e.g. the cyto- or myeloarchitectonic techniques), is that it does not allow one to infer specific functions for the areas that it maps. Paradoxically, this may also be its main strength since it does not depend upon the simplistic and sometimes erroneous attribution of a single function. Nevertheless, a reverse correlation of a region's ATC with the stimulus may generate a hypothesis with regard to its preferred feature (Hasson et al. 2004), which should then be validated statistically using a formal hypothesis-led regression against the intensity of the feature under concern (Bartels & Zeki 2003). We have demonstrated that a multitude of distinct regions can be identified across the whole brain, even within the visual cortex, based entirely on their differential ATCs during natural viewing. Functionally corresponding regions did not only correspond anatomically across subjects, but also had highly area-specific temporal inter subject correlations. In contrast, distinct functional subdivisions, even within visual cortex, had surprisingly low temporal correlations, even within a single brain. In fact, our results show that natural conditions lead to a decorrelation of activity between distinct regions compared with rest (Bartels & Zeki 2004a, 2005). This makes it easier for the ICA to identify voxels belonging to a given subdivision and to segregate them into separate components. At the same time, it also makes functionally connected (i.e. temporally correlated) regions stand out. In fact, our results also showed that regions with known substantial anatomical connections increased the correlations of their ATCs during natural viewing compared with the rest, revealing a double dissociation as a function of anatomical connectivity and stimulation condition. Temporal correlations during natural conditions may thus be indicative of anatomical connectivity, which we demonstrate for the case of regions involved in the processing of language. Most importantly, our results reveal a high degree of temporal independence of distinct cortical regions during natural processing, which not only confirms the view of a high degree of functional specialization even within the visual cortex, but also indicates that natural conditions may enhance regional specificity (Zeki 1978; Zeki & Bartels 1998b). Other types of natural stimuli may be more suitable to obtain differential activity in other parts of the brain, and future algorithms may deliver better results than the ICA used here. However, we hope that the general principle of chronoarchitectonic mapping may serve to identify novel cortical subdivisions and their connectivity. The results obtained may then inspire future, hypothesis-led experiments to illuminate the precise functional involvement of these regions.