Brain Network Analysis (using Atlases) Primer

Date: January 19, 2024 2:41 PM

Brain functional-connectivity network-analysis data pipeline is generally:

voxel → region → network.

So the voxels aren’t a network, the voxels get assigned to regions, and then network analysis is done on the 100-500 regions (which are groups of the voxels), with each of the regions being one ‘node’ in the brain network.

Brain Network Analysis Pipeline

First level of analysis: Voxels from the fMRI scan directly

  • ex: cortex surface = ~60,000 vertices
    • do alignment, corrections, other preprocessing
  • 2 common preprocessing pipelines:
    • Glasser 2013, aka Human Connectome Project Pipeline
    • fMRI Prep
  • lots of distortion corrections to be done to find the same voxel in two different brains, if you’re trying to cross-examine voxels

Second level: the region-level

  • Algorithms tend to find 100-500 regions
  • Finding communities, so that you can run the analysis on 100-500 nodes, instead of 60,000 nodes (which would take years)
  • You can use an algorithm to find the communities/modules in your data
    • looking for voxels with similar patterns of activation (aka “coactivation”, “time-series dependencies”, “functional connectivity”)
    • Using methods like louvin communities, or machine learning methods, or clustering methods, community detection, multi-modal parcellation, clustering, physical location, historical precedent
      • this is the parcellation
      • brain regions are referred to as brain parcels
      • tho people may also say ‘brain node’
  • Can also apply other researchers’ atlases to your fMRI data, instead of finding your own regions
    • called “partitioning” or “parcellation” or “using a hard partition”
    • list of all atlases: https://www.lead-dbs.org/helpsupport/knowledge-base/atlasesresources/cortical-atlas-parcellations-mni-space/
    • if you were to make a adjacency matrix of the FC of all your voxels to each other, you’d see a random array of correlations
      • partition (by an atlas) = indexing vector to sort the matrix
        • shows the communities structure
    • take the vertices/voxels from your data and then use someone else’s region parcellation (like yeo, shafer, glasser) and it indexes which vertices → which region
      • 5 total popular partitions
      • Partitions used because people kept applying community detection and found the same networks/communities/modules over and over again
      • this is called using a ‘hard partition’
      • other people have found these partitions

Third level: brain network

  • take regions and connect them into networks - based on coactivation (activity measuring -> Pearson’s correlation) for all of the regions against each other
  • divide those regions into groups to find 5-20 networks
    • network = statistical dependency, pattern of coactivation
      • nodes = physical regions defined by the positions of the voxels (or verticies, if you’re using surface data) on the brain which were indexed into that region
    • edge value between two nodes = statistical dependency of time series between two regions
      • pearsons correlation of activation is a common method to calculate statistical dependency
  • people may say ‘partition’ for network
  • or may say ‘parcellation’ for network
  • or ‘atlas’”

Common Atlases

  • Yeo, 2011 atlas (aka ‘Yeo regions’ aka ‘400 yeo/shafer regions’)
  • Glasser, 2013 atlas (aka MMP aka ‘multi-modal parcellation’)
    • Using another researcher’s found partition is called “using a hard partition”

What is the scale of this data?

Vertex

  • 160,000 vertices across a brain = the mesh
  • ‘vectorized topology’
  • taken down to 91,000 number of usable vertices
    • cortex/cortical vertices: 60,000
    • subcortical: 31,000

Functional Connectivity Matrices

  • usually at the region level
    • region-level analysis
  • since using 60,0000x60,0000 fc matrix would take forever to do a whole dataset
    • or you can choose specific regions or other ways to reduce the # of vertices
  • MVPA or RSA still use voxel- or vertex-level analysis
    • what pattern of activation is most related to some function/task
  • There’s also first- and second-level analyses aka task contrast
    • This is for quality control
    • To relate a specific region with a task
    • That will use voxel- or vertex-level analysis
      • Statistical significance testing
      • Multiple comparison testing
  • Parcellation: Brain Region Level
    • 100-500 regions
      • people may also say ‘brain node’, ‘node’, ‘parcel’, ‘brain parcel’ instead of ‘region’
    • this is the ‘parcellation’
  • Brain networks
    • they take regions and connect them into networks
    • 5-20 networks
    • People may say ‘partition’ for network
    • or may say ‘parcellation’ for network
    • or ‘atlas’