Construction and Evaluation of Hierarchical Parcellation of the Brain using fMRI with Prewhitening

TL;DR

This paper develops and compares hierarchical brain parcellations derived from fMRI data with and without prewhitening, evaluating their consistency, homogeneity, and reproducibility to improve functional brain mapping.

Contribution

It provides a rigorous evaluation of hierarchical clustering for brain parcellation and explores the impact of prewhitening on the resulting atlas structure.

Findings

Prewhitening improves the homogeneity of parcellations.

Reproducibility across subjects is higher with prewhitened data.

Hierarchical clustering produces consistent brain regions aligned with known anatomy.

Abstract

Brain atlases are a ubiquitous tool used for analyzing and interpreting brain imaging datasets. Traditionally, brain atlases divided the brain into regions separated by anatomical landmarks. In the last decade, several attempts have been made to parcellate the brain into regions with distinct functional activity using fMRI. To construct a brain atlas using fMRI, data driven algorithms are used to group voxels with similar functional activity together to form regions. Hierarchical clustering is one parcellation method that has been used for functional parcellation of the brain, resulting in parcellations that align well with cytoarchitectonic divisions of the brain. However, few rigorous data driven evaluations of the method have been performed. Moreover, the effect of removing autocorrelation trends from fMRI time series (prewhitening) on the structure of the resultant atlas has not been previously explored. In this paper, we use hierarchical clustering to produce functional parcellations of the brain using hierarchical clustering. We use both prewhitened and raw fMRI time series to construct the atlas. The resultant atlases were then evaluated for their homogeneity, separation between regions, reproducibility across subjects, and reproducibility across scans.