Data-driven clustering reveals a fundamental subdivision of the human cortex into two global systems

Yulia Golland, Polina Golland, Shlomo Bentin, Rafael MalachNeuropsychologia 2007 · DOI

TL;DR

This paper uses an unsupervised clustering approach on fMRI data to reveal a fundamental, two-system organization of the human cortex. This data-driven finding strongly supports previous hypotheses about a global cortical division into "extrinsic" (sensory/motor) and "intrinsic" (internally-oriented) networks, offering a robust and consistent anatomical framework for understanding brain function.

What Problem Does This Paper Address?

Understanding how the human brain is organized at a global scale is crucial for deciphering its complex functions. Previous research has proposed large-scale cortical partitions, such as modality-specific systems or major organizational principles. A recent study by the same authors proposed a "intrinsic-extrinsic" subdivision of the posterior cortex, identifying an "extrinsic" system linked to external sensory input and an "intrinsic" system associated with internally-driven processing, overlapping with the "default-mode network." However, these prior findings were based on "seed-based" functional connectivity analyses, which are limited by the need for pre-determined hypotheses, the selection of seed regions, and arbitrary correlation thresholds. This paper aims to overcome these limitations by applying a hypothesis-free, unsupervised clustering algorithm to a large, diverse fMRI dataset, seeking to confirm and robustly establish this fundamental bipartite cortical division.

Methods

The core methodological innovation is the application of an unsupervised, hypothesis-free k-means clustering algorithm to a large fMRI dataset. This algorithm assigns voxels to a predefined number of clusters (here, primarily k=2) based on the similarity of their time-courses, eliminating the need for pre-selected "seed" regions or correlation thresholds. The algorithm iteratively estimates cluster membership and updates mean time courses. To define similarity between time courses, the authors used a correlation coefficient, and the squared distance was defined as 1 minus this correlation. This was extended to handle multiple experiments by averaging across them. To ensure robustness, they also evaluated Gaussian mixture modeling and spectral clustering, finding highly similar results. For group analysis, clustering was performed individually, and then voxel labels were transformed into Talairach space to assess consistency across participants, creating "group consistency maps." Activation/deactivation maps were generated using the General Linear Model to compare with the clustering results.

Datasets

The study utilized a large dataset of fMRI experiments from seven participants, totaling 372 minutes of EPI scan time. These experiments covered a variety of functional characteristics, including:

  • Repeated movie experiment (audio-visual movie viewing)
  • Rest experiment (eyes closed)
  • Visual localizer (block-design visual experiment)
  • Face morphing experiment
  • Internal tasks experiment (mental calculation, internal speech, semantic memory)
  • Protocol-driven visual experiments (used for activation/deactivation analysis)

The paper does not introduce new benchmark datasets but rather aggregates existing data from multiple experiments to provide a more comprehensive and robust analysis.

Key Insights and Ideas

  • Unsupervised Clustering Confirms Global Cortical Division: The most significant insight is that a purely data-driven, hypothesis-free clustering approach robustly identifies a fundamental bipartite division of the human cortex, mirroring the previously hypothesized "intrinsic" and "extrinsic" systems. This suggests the division is not an artifact of specific experimental designs or analytical choices.
  • Intrinsic Network's Dissociation from External Input: The "intrinsic" network (largely overlapping with the default-mode network) shows significantly higher inter-subject variability in its activity during external stimulation (visual tasks, movie viewing) compared to the "extrinsic" network. This supports the idea that its activity is less driven by external sensory input and more by internal processes.
  • Extrinsic Network's Role in External Processing: The "extrinsic" network, associated with sensory and motor areas, exhibits more consistent activity across subjects during external stimulation, aligning with its presumed role in processing external information.
  • Robustness Across Methods and Participants: The consistency of the clustering results across different participants and even across different clustering algorithms (k-means, Gaussian mixture, spectral clustering) strongly suggests the identified partition is a fundamental property of cortical organization.
  • Hierarchical Organization Potential: The exploration of k=3 clustering revealed that the extrinsic system can be further subdivided, suggesting a hierarchical organization of cortical function beyond the initial bipartite split.

Key Results

  • The unsupervised k-means clustering algorithm consistently delineated two major cortical systems across all participants.
  • The anatomical topography of these two systems closely matched the previously defined "extrinsic" (sensory-motor, associated with external environment) and "intrinsic" (overlapping with default-mode network, associated with internal processing) systems.
  • Inter-subject variability of activity was significantly lower in the extrinsic system than in the intrinsic system during sensory stimulation (visual localizer: extrinsic 0.57±0.07 vs. intrinsic 0.77±0.02; movie: extrinsic 0.73±0.01 vs. intrinsic 0.87±0.02).
  • During rest, the inter-subject variability of activity in both systems was similar.
  • The anatomical boundaries of the intrinsic system identified by clustering showed strong agreement with those reported in previous studies of the default-mode network.
  • When exploring k=3 clustering, the extrinsic system was consistently subdivided into visual areas and other sensory/motor/attentional areas.

Limitations and Future Work

  • Neuroanatomical Variability: The authors acknowledge that while the clustering reveals consistent global patterns, individual anatomical differences exist. They suggest that advancements in spatial normalization methods could further improve group-wise comparisons. They also note that some prefrontal regions might have intermediate functionality, not fitting neatly into either system, and suggest further research with paradigms emphasizing frontal activations.
  • Functional Specificity of Intrinsic System: The precise functional role of the intrinsic system remains enigmatic. While its decoupling from external input is a key characteristic, its exact contribution to cognition requires further elucidation. The term "intrinsic" is proposed to refer to its self-directed orientation rather than solely spontaneous activity.
  • Hierarchical Exploration: The paper briefly explores k=3 and k=4 clustering, suggesting that the identified systems can be further subdivided. Future work could systematically explore these finer-grained hierarchical partitions to map out more functionally specific sub-systems.
  • Data-Driven vs. Hypothesis-Driven: While the study demonstrates the power of data-driven approaches, it also acknowledges the value of hypothesis-driven research in refining our understanding of specific brain regions and their roles.

Summary updated

BibTeX
@inproceedings{golland2007data,
  author = {Yulia Golland and Polina Golland and Shlomo Bentin and Rafael Malach},
  title = {{Data-driven clustering reveals a fundamental subdivision of the human cortex into two global systems}},
  booktitle = {Neuropsychologia},
  year = {2007},
  doi = {10.1016/j.neuropsychologia.2007.10.003},
}
Figure