Abrupt spatial changes in anatomic and functional properties of the brain demarcate boundaries between discrete functional areas. While previous work has identified these boundaries in cortex, a new study by Tian et al. applies this approach for the first time to subcortical structures within the in vivo human brain.
Understanding the organization of the human brain is a critical goal of systems neuroscience. The best evidence available from humans and nonhuman primates suggests that cortex, at least, is organized into a large number of discrete cortical areas1,2 with sharp boundaries between areas at which multiple brain properties, including function, architectonics, connectivity and/or topographical patterns, abruptly change3. In principle, mapping human brain areas by identifying the boundaries between them can also be accomplished and is the mapping approach most convergent with a body of invasive animal work that has provided substantial insight into brain organization. Indeed, previous work has made substantial progress in noninvasively mapping such areal transitions in the in vivo human cortex and thus putatively delineating (‘parcellating’) the areas themselves, using functional MRI (fMRI) resting-state functional connectivity (RSFC) approaches4,5,6,7.
However, by contrast with the cortex, relatively few attempts have been made to comprehensively parcellate human subcortex using RSFC8,9, and no previous work has attempted to do so by identifying boundaries between discrete regions. This is likely in part because of the additional technical challenges inherent in identifying such borders in subcortical structures rather than in the cortical sheet. Subcortical structures are farther from the MRI coil than cortex, reducing data quality. Subcortical structures are also smaller relative to fMRI data resolution, making it more difficult to observe sharp transitions in data that has a given spatial resolution. Then too, potential transitions between areas are more complex to describe in a three-dimensional object than on the two-dimensional cortical sheet. Finally, it is not as well established that transitions between regions are always abrupt in subcortex. Some subcortical transitions may be gradual, due to the complex integration of inputs from multiple sources10.
In this issue of Nature Neuroscience, Tian et al. introduce a ‘gradientography’ approach that aims to address these issues11. A key insight of this approach is that shifting RSFC properties can be envisioned as directional, moving continuously through a volume (Fig. 1). Thus, borrowing concepts from established diffusion MRI tractography approaches, the authors estimate the RSFC gradient for each voxel as a tensor, which represents the magnitude and orientation of the major change in RSFC. From a single voxel’s tensor, they then propagate gradient streamlines running from voxel to voxel along the direction of RSFC change, and thus trace the paths such changes follow throughout the subcortical volume (Fig. 1). These gradient streamlines are not restricted to be unidirectional and linear, but may follow any complex curvature representing the spatial organization of RSFC changes in subcortex. This revolutionary technique allows the authors to identify transitions in RSFC patterns that are not just purely local, as in current cortical boundary mapping approaches5,6, but that continuously span large segments of a subcortical structure. Locations where such transitions are particularly sharp (Fig. 1) are likely to represent borders between discrete functional regions in subcortex.
Critically, the authors also describe how apparent increases in gradient magnitude, which could be interpreted as transitions between regions, may be driven not only by veridical transitions in RSFC, but also by confounding effects such as the geometry of the subcortex. This is a very important observation that was not intuitively obvious, but can significantly confound attempts to identify subcortical boundaries. This knowledge allowed the authors to not only identify subcortical locations with large transitions in RSFC, but also to compare them to a null model controlling for subcortical geometry.
The resulting null-corrected peaks in subcortical transitions form the basis for a comprehensive, multiscale hierarchical parcellation of striatum, thalamus, hippocampus and amygdala. Excitingly, multiple boundaries in this parcellation recapitulate known divisions between anatomical structures, which has not always been the case for previously attempted parcellations of subcortex. One important example here is the division between the anterior hippocampus and the amygdala. While these adjacent structures are consistently separable based on anatomical scans using automated segmentation software, functional parcellations often observe few clear differences between the connectivity patterns of the anterior hippocampus and the amygdala8,9. The ability to recapitulate such known anatomical divisions that exhibit only very subtle functional differences is a major technical advance.
At the most divided scale, this parcellation moves substantially beyond known anatomical divisions to identify divisions within structures that appear homogenous to an anatomical MRI scan, but that are also consistent with known functional and cytoarchitectonic divisions. For example, the parcellation distinguishes nucleus accumbens core from shell; it distinguishes caudate head from body from tail; and it distinguishes hippocampus tail from body from head, as well as medial and lateral divisions within hippocampus head. Accurate delineation of subdivisions within these structures is the first step toward understanding how these different subdivisions may be contributing to human cognition.
An important consideration for the practical use of any parcellation atlas is how well it can be applied to new populations and individuals of interest. Applying such atlases to new individuals ideally requires creating individual-specific versions of the atlas, due to the known cross-individual variability in brain organization6,12. Here the authors train a classifier to identify RSFC signatures of each parcel, which can be used to label each voxel in a new individual as belonging to a specific parcel, thus generating participant-specific versions of each parcel. If such classifications can be trusted, then the resulting individual-specific parcellations represent an ideal functional registration between individuals, allowing direct cross-participant comparisons of anatomy, activity and/or connectivity, not confounded by interindividual spatial variability.
Validation of these classifications suggests that they are indeed highly successful in most cases, which is very promising. However, a minority of participants exhibited markedly poor correspondence between their estimated individual-specific parcellation and the atlas. It is possible that these are people who exhibit unusual, idiosyncratic organization of their subcortical structures—similar to the idiosyncratic organization of cortical networks that has previously been described—which would be of notable interest for further investigation to determine if such idiosyncratic organization is relevant to behavior or cognition. On the other hand, it is also possible that these low-corresponding participants represent individuals with poor data quality, precluding accurate voxel classification, or simply represent failures of the classifier. In these latter cases, the participants should not be included or investigated as if their parcellations are valid. Future work in this area should focus on determining whether such participants with unusual individual-specific parcellations are reasonable topics of study or should be excluded from future work. Finally, it should be noted that the present classifier must be further tested to ensure that it works effectively on non-Human Connectome Project datasets, as the HCP is in many ways a highly idiosyncratic dataset, and classifiers can often struggle to generalize to datasets collected using scanners and sequences they have not been trained on.
Perhaps the most interesting finding in this work was that the performance of cognitive tasks shifted some parcellation boundaries relative to their position during the resting state. For example, task performance appeared to eliminate the observed boundary between amygdala and anterior hippocampus, as well as between nucleus accumbens and caudate or putamen. This observation follows a similar recent report of shifted parcel borders driven by task performance in cortex13. The authors interpret this finding as reflecting adaptive recruitment of sub-parcel neural populations by different brain regions in response to changing task demands. This is a fascinating possibility. However, it does raise the critical question of what these parcellations actually represent. The original conceptualization of discrete brain areas being separated by sharp boundaries1,2 does not allow for the dynamic reconfiguration of areas defined in part by exhibiting differential architectonics, topography and anatomical connectivity, as these factors should not be influenced by varying task contexts. While it is plausible that under certain task conditions, amygdala and anterior hippocampus may work in close concert and thus exhibit similar observable functional responses, this does not necessarily mean that we should consider these two structures as dynamically forming a unified brain parcel. Future work must establish (i) whether the hypothesis of dynamic recruitment of subareal neural populations can be supported using non-fMRI methods and (ii) under what task conditions such fMRI-derived parcellations best represent actual anatomically supported divisions of the brain. The present work by Tian et al. provides an exceptionally strong starting point with which to pursue these questions.
Sejnowski, T.J. & Churchland, P.S. Brain and cognition. in Foundations of Cognitive Science (ed. Posner, M. I.) 888 (MIT Press, 1989).
Brodmann, K. Vergleichende Lokalisationslehre der Grosshirnrinde in ihren Prinzipien dargestellt auf Grund des Zellenbaues (Barth, 1909).
Felleman, D. J. & Van Essen, D. C. Cereb. Cortex 1, 1–47 (1991).
Wig, G. S., Laumann, T. O. & Petersen, S. E. Neuroimage 93, 276–291 (2014).
Gordon, E. M. et al. Cereb. Cortex 26, 288–303 (2016).
Glasser, M. F. et al. Nature 536, 171–178 (2016).
Schaefer, A. et al. Cereb. Cortex 28, 3095–3114 (2018).
Ji, J. L. et al. Neuroimage 185, 35–57 (2019).
Seitzman, B. A. et al. Neuroimage 206, 116290 (2019).
Haber, S. N. J. Chem. Neuroanat. 26, 317–330 (2003).
Tian, Y., Margulies, D.S., Breakspear, M. & Zalesky, A. Nat. Neurosci. https://doi.org/s41593-020-00711-6 (2020).
Gordon, E. M. et al. Neuron 95, 791–807.e7 (2017).
Salehi, M. et al. Neuroimage 208, 116366 (2020).
The author declares no competing interests.
About this article
Cite this article
Gordon, E.M. Functional boundaries within human subcortex. Nat Neurosci 23, 1312–1314 (2020). https://doi.org/10.1038/s41593-020-00721-4