Functional parcellation using time courses of instantaneous connectivity

Erik S.b. van Oort, Maarten Mennes, Tobias Navarro Schröder, Vinod Kumar, Néstor Jiménez, Wolfgang Grodd, Christian F. Doeller, Christian F. BeckmannNeuroImage 2017 · DOI

TL;DR

This paper introduces a novel "top-down" method called Instantaneous Connectivity Parcellation (ICP) to divide brain regions into functionally distinct sub-regions. ICP leverages subtle temporal dynamics in fMRI data to achieve biologically plausible parcellations that align well with known brain anatomy, offering a more refined understanding of brain organization.

What Problem Does This Paper Address?

Understanding the brain's functional architecture relies on identifying distinct, functionally specialized sub-regions within larger brain areas. Current methods often employ a "bottom-up" approach, starting with small units and grouping them into larger networks. However, this can be limited by the initial definition of these small units, which may not always reflect true biological boundaries and can propagate inaccuracies. Furthermore, accurately parcellating smaller, deeper brain structures with high functional differentiation is challenging due to limitations in MRI signal sensitivity. This paper addresses these limitations by proposing a "top-down" strategy that starts with a larger region of interest and refines it into sub-regions based on instantaneous connectivity dynamics.

Methods

The core of the proposed method is Instantaneous Connectivity Parcellation (ICP). It works by:

  1. Defining an initial Region of Interest (ROI): This can be a large brain area or even a whole-brain mask.
  2. Extracting the average time course of the ROI.
  3. Transforming the fMRI time series of individual voxels within the ROI to capture "instantaneous" connectivity. This is achieved by calculating the time-resolved Pearson correlation between each voxel's time course and the ROI's average time course. This process amplifies subtle temporal differences and transient events, making sub-regions more distinguishable.
  4. Applying Group Independent Component Analysis (GICA) to the transformed data to identify spatially distinct sub-regions.
  5. Determining the optimal number of sub-regions using split-half reproducibility. This involves splitting the data into halves, generating parcellations for each half, and calculating the Dice overlap between them across a range of scales (number of sub-regions). The scale with the highest reproducibility is chosen.

This approach is contrasted with traditional "bottom-up" methods like clustering and standard ICA, which rely on temporal averages and can struggle with subtle differences or smaller structures.

Datasets

The authors utilized two primary datasets:

  • Human Connectome Project (HCP): Data from 100 subjects, featuring high-quality resting-state fMRI (R-FMRI) with a rapid TR (720ms) and advanced denoising (FSL-FIX). This dataset was used for parcellating sensory-motor cortex and subcortical structures.
  • 7T fMRI data: Acquired locally at the Donders Institute from participants performing a spatial navigation task. This high-resolution data (0.9mm isotropic voxels) was used specifically for parcellating the entorhinal cortex.

Key Insights and Ideas

  • Top-Down Refinement: ICP offers a principled way to move from large functional regions to finer sub-regions, leveraging temporal dynamics rather than just spatial proximity or static correlations.
  • Sensitivity to Temporal Dynamics: By focusing on "instantaneous" connectivity, ICP can detect subtle, transient events that are averaged out in traditional methods, leading to more precise parcellations.
  • Biologically Valid Subdivisions: The parcellations generated by ICP show strong correspondence with known cytoarchitectonic atlases, suggesting that the method captures biologically meaningful organizational principles.
  • Hierarchical Parcellation Potential: The method naturally lends itself to hierarchical parcellation, as seen in the subcortical results where coarser scales reveal larger structures that subdivide into finer ones at higher scales.
  • Addressing Small, Deep Structures: ICP demonstrates promise in parcellating challenging regions like the thalamus and entorhinal cortex, which are difficult for surface-based methods or those with lower spatial resolution.

Key Results

The ICP method was successfully applied to several brain regions, demonstrating its utility:

  • Sensory-Motor System: ICP yielded a 12-sub-region parcellation that reflected both anterior-posterior organization (matching the Jülich atlas) and the somatotopic (homunculus) organization along the central sulcus. It showed good overlap with the Jülich atlas, outperforming a Craddock NCUT parcellation.
  • Subcortical Structures: ICP produced both coarse (8 sub-regions) and fine-grained (27 sub-regions) parcellations of the brainstem, striatum, and hippocampus. The fine-grained parcellation showed high overlap with the Harvard-Oxford atlas, significantly outperforming ICA and Craddock parcellations (e.g., average Dice of 0.67 for ICP vs. 0.54 for ICA and 0.29 for Craddock for the brainstem).
  • Thalamus: A 30-sub-region parcellation of the thalamus showed excellent overlap with the Morel histological atlas for several nuclei (e.g., Red Nucleus: Dice 0.96). ICP significantly outperformed ICA and Craddock parcellations in thalamic parcellation.
  • Entorhinal Cortex: ICP generated a 9-region parcellation of the entorhinal cortex that qualitatively aligned with known cytoarchitectonic subdivisions, highlighting its potential for mapping complex, small structures.

Limitations and Future Work

  • Split-Half Reproducibility Caveats: The authors acknowledge that split-half reproducibility can be biased by the Dice overlap calculation, which favors larger regions. They suggest selecting scales based on local maxima in reproducibility. Additionally, peaks at multiple scales can occur, representing different hierarchical levels, and the choice of scale depends on the research question.
  • SNR Dependency: The reliability of parcellations is dependent on data quality and quantity, with Signal-to-Noise Ratio (SNR) being a limiting factor. The use of large, high-quality datasets like HCP is crucial.
  • Future Work: The authors plan to extend ICP to generate a hierarchical, full-brain parcellation. They also aim to investigate its application in clinical populations and for studying specific cognitive processes.

Summary updated

BibTeX
@inproceedings{oort2017functional,
  author = {Erik S.B. van Oort and Maarten Mennes and Tobias Navarro Schröder and Vinod Kumar and Néstor Jiménez and Wolfgang Grodd and Christian F. Doeller and Christian F. Beckmann},
  title = {{Functional parcellation using time courses of instantaneous connectivity}},
  booktitle = {NeuroImage},
  year = {2017},
  doi = {10.1016/j.neuroimage.2017.07.027},
}
Figure