Functional architecture of cerebral cortex during naturalistic movie-watching

Reza Rajimehr, Haoran Xu, Asa Farahani, Simon Kornblith, John Duncan, Robert DesimonebioRxiv 2022 · DOI

TL;DR

This paper presents a novel functional parcellation of the entire human cerebral cortex by analyzing fMRI data from subjects watching a 60-minute movie. This data-driven approach reveals 24 distinct functional areas/networks, including three executive control networks that exhibit a "push-pull" interaction with domain-specific regions, offering a more comprehensive and unified map than traditional methods.

What Problem Does This Paper Address?

Understanding how the brain processes information relies heavily on characterizing the functional organization of the cerebral cortex. While previous neuroimaging studies have identified the functions of many cortical areas using specific stimulus categories, it remains unclear how these areas are organized during more naturalistic, complex stimuli like movies. Traditional methods often rely on "functional localizer" scans, which are time-consuming and may not capture the full functional repertoire of an area. This paper aims to address this gap by creating a comprehensive functional map of the entire cortex using a single, rich audiovisual stimulus, providing a unified view of brain organization during naturalistic engagement.

Methods

The researchers employed a data-driven clustering approach on high-resolution fMRI data.

  1. Data Preprocessing: fMRI data from individual subjects were preprocessed and registered to a standard cortical surface. Time-courses of activity were de-meaned and concatenated across functional runs, then averaged across subjects to capture inter-subject synchronization during natural vision.
  2. Activity Space Construction: An "activity space" was created where each axis represented the functional activity at a specific time-point. Vertices of the cortical surface were treated as data points in this high-dimensional space.
  3. Hierarchical Clustering: An agglomerative hierarchical clustering algorithm was used to group vertices based on their geometric distance in the activity space. This method allows for partitioning at various scales, from coarse to fine.
  4. Reliability Assessment: The similarity of clustering maps between the left and right hemispheres was assessed using Fowlkes-Mallows and adjusted Rand indices. Control analyses with simulated data confirmed the reliability of the clustering for real data.
  5. Parcellation Selection: A cutoff of 24 clusters was chosen based on the stability of inter-hemispheric similarity at higher cluster numbers.
  6. Analysis of Functional Properties: The paper further analyzed response variability within clusters and the correlation between cluster activity and eye-movement variability. They also investigated the relationship between their parcellation and established functional localizer maps and networks.

Datasets

  • Human Connectome Project (HCP) 7T dataset: This dataset, released in April 2018, provided the primary fMRI data for 176 healthy young adults (aged 22-35) who underwent movie-watching and resting-state scans.
  • Action Dataset: Used for an independent action localizer experiment, this dataset contains 300 video clips from 5 action categories.
  • HCP working memory task data: Used for functional localizer maps for visual categories (faces, bodies, tools, places).
  • HCP language processing task data: Used for functional localizer maps for auditory-language networks.
  • HCP social cognition task data: Used for functional localizer maps for social cognition networks.
  • Yeo's 7-network parcellation: Used as a reference for identifying default mode and dorsal attention networks.

Key Insights and Ideas

  • Naturalistic stimuli reveal robust cortical organization: Watching a movie, a complex and dynamic stimulus, is sufficient to drive and parcellate the entire cerebral cortex into functionally coherent regions. This challenges the necessity of highly specific, artificial stimuli for mapping brain function.
  • "Push-pull" interaction between domain-general and domain-specific networks: The study identified three executive control (domain-general) networks that show a strong negative correlation (push-pull) with domain-specific regions (e.g., visual, auditory, language). This suggests an efficient neural resource allocation where domain-general systems are recruited when stimuli are ambiguous or require effortful resolution, while domain-specific systems handle clear content.
  • Hierarchical clustering reveals multi-scale organization: The use of hierarchical clustering allowed for the identification of functional areas at different levels of granularity, reflecting the nested nature of cortical organization.
  • Movie-watching data can replace traditional localizers: The comprehensive map generated from movie-watching data could potentially replace the need for numerous, time-consuming functional localizer scans for many research questions.
  • Transient executive control network activity at movie transitions: A surprising finding was a significant, transient increase in activity in executive control networks at the transition from movie to rest periods, suggesting a role in memory consolidation or internal model disassembling.

Key Results

  • A data-driven clustering approach on 60 minutes of movie-watching fMRI data yielded a functional parcellation of the cerebral cortex into 24 distinct areas/networks.
  • These 24 clusters showed high similarity between the left and right hemispheres, indicating reliable functional organization.
  • The parcellation revealed three executive control (domain-general) networks that exhibited a significant "push-pull" interaction with domain-specific regions.
  • Category-selective clusters (e.g., face, body, object/tool, scene) from the movie data showed strong correspondence with regions identified by traditional category-selective localizer tasks.
  • The action perception network showed selective responses to dynamic human-object interactions.
  • Executive control networks showed a transient, significant response at the end of movie clips, not observed at the beginning.

Limitations and Future Work

  • Group-average data: The parcellation was based on group-average time-courses, which might obscure idiosyncratic parcellation maps present in individual subjects. Future work could focus on single-subject parcellation to reveal finer details and individual differences.
  • "Hard borders" in clustering: The clustering algorithm imposes hard borders between regions, while actual functional selectivity might be more continuous. Investigating regions at cluster borders with preferred movie frames could clarify this.
  • Stimulus correlations: The naturalistic movie stimulus contains highly correlated information (e.g., faces and bodies often appear together). While the clustering algorithm could separate some of these areas, it highlights the challenge of disentangling co-occurring stimuli.
  • Future Directions: The authors suggest that this parcellation scheme can be used to investigate functional interactions between areas during movie-watching, test hypotheses about cluster functions using reverse-correlation with movie frames, and be applied to study cortical organization in disease populations and during lifespan development. The "push-pull" interaction also warrants further investigation with controlled paradigms and deep neural networks.

Summary updated

Abstract

Abstract Characterizing the functional organization of cerebral cortex is a fundamental step in understanding how different kinds of information are processed in the brain. Neuroimaging studies over the past decades have uncovered the function of many cortical areas in response to selected stimulus categories. However, it is still unclear how cortical areas are organized during naturalistic visual and auditory stimulation. Here we used high-resolution functional MRI data from 176 human subjects to map the macro-architecture of the entire cerebral cortex based on responses to a 60-minute audiovisual movie stimulus. A data-driven clustering approach revealed a map of 24 functional areas/networks, each related to a specific aspect of sensory or cognitive processing. The map included three distinct executive control (domain-general) networks which showed a strong push-pull interaction with domain-specific regions in visual, auditory, and language cortex. The cortical parcellation scheme presented here provides a comprehensive and unified map of functionally defined areas, which could replace a large set of functional localizer maps.
BibTeX
@misc{rajimehr2022functional,
  author = {Reza Rajimehr and Haoran Xu and Asa Farahani and Simon Kornblith and John Duncan and Robert Desimone},
  title = {{Functional architecture of cerebral cortex during naturalistic movie-watching}},
  year = {2022},
  doi = {10.1101/2022.03.14.483878},
  howpublished = {bioRxiv},
}
Figure