Evaluation of full brain parcellation schemes using the NeuroVault database of statistical maps

EVALUATION OF FULL BRAIN PARCELLATION SCHEMES

USING THE NEUROVAULT DATABASE OF STATISTICAL MAPS

Krzysztof J. Gorgolewski, Arielle Tambini, Joke Durnez, Vanessa V. Sochat, Joe Wexler, Russell A. Poldrack

Motivation•There are many different ways to divide the brain into regions.

•What makes one parcellation better than other?

Evaluate different parcellation scheme using

task fMRI activation patterns

Goal

Methods: Included atlases

Name Number of parcels

Coverage Properties

Gordon et al. 333 cortical asymmetricalAAL 116 cortical +

subcorticalasymmetrical

Collins et al. 9 cortical + subcortical

symmetrical

Yeo et al. 7 or 17 cortical symmetrical networks

Glasser et al. 180 or 360 cortical symmetrical or asymmetrical

Harvard-Oxford 58 cortical + subcortical

symmetrical

Brainnetome 246 cortical + subcortical

asymmetrical

Methods: Metrics• A good parcellation should accurately delineate activation patterns from task fMRI

• Variance of T/Z values within parcels should be minimized

• Homogenous brain areas should be covered by minimal number of parcels

Methods: Procedure1. Gather a large collection of statistical

maps2. For each map:

1. Calculate mean within parcel variance2. Calculate between parcel variance3. Between/within parcel variance ratio

should be higher for better parcellations

Methods: Statistical maps• Inclusion criteria:

• From a published study• In MNI space• Unthresholded• Task fMRI

Final collection:• 625 statistical maps from 79 papers• Representing 87 different tasks

Methods: Included statistical maps

Methods: Confounds and how to deal with themConfounds:1. Parcellation with more regions have smaller parcels

and thus lower within parcel variance2. Each input map has different smoothness

Instead of asking: How well does this parcellation represent this activation pattern?

Ask:How this parcellation represents this activation pattern in contrast to a random pattern?

Methods: Null distributionFor each statistical map:1. Estimate smoothness 2. Shuffle values3. Smooth4. Repeat x1000

Example statistical map

Randomized statistical map with matched smoothness

Within parcel variance

big parcels small parcels

Normalizationbefore after

Normalized between/within parcel variance

Top symmetricalName Between/within parcel

variance [z-score]Within parcel variance[z-score]

Yeo et al. (17) 25.84 -5.43

Yeo et al. (7) 24.19 -2.77

Glasser et al. (180) 22.62 -7.89

Harvard-Oxford 14.80 -5.04

Top asymmetricalName Between/within parcel

variance [z-score]Within parcel variance [z-score]

Glasser et al. (360) 14.91 -6.82

Shen et al. (100) 14.34 -8.40

Brainnetome 13.92 -7.38

Shen et al. (50) 13.88 -7.37

Conclusions• Most parcellations perform well above chance

• Symmetrical parcellation do better

• Yeo networks perform particularly well

• There is no clear winner

Future directions• Surface based comparison• Subcortical only comparison• Improve manual quality control