Atlas-Based Labeling of Resting-State fMRI

Abstract

Background:

Functional magnetic resonance imaging (fMRI) has the potential to provide noninvasive functional mapping of the brain with high spatial and temporal resolution. However, fMRI independent components (ICs) must be manually inspected, selected, and interpreted, requiring time and expertise. We propose a novel approach for automated labeling of fMRI ICs by establishing their characteristic spatio-functional relationship.

Methods:

The approach identifies 9 resting-state networks and 45 ICs and generates a functional activation feature map that quantifies the spatial distribution, relative to an anatomical labeled atlas, of the z-scores of each IC across a cohort of 176 subjects. The cosine-similarity metric was used to classify unlabeled ICs based on the similarity to the spatial distribution of activation with the pregenerated feature map. The approach was tested on three fMRI datasets from the 1000 functional connectome projects, consisting of 280 subjects, that were not included in feature map generation.

Results:

The results demonstrate the effectiveness of the approach in classifying ICs based on their spatial features with an accuracy of better than 95%.

Conclusions:

The approach significantly reduces expert time and computation time required for labeling ICs, while improving reliability and accuracy. The spatio-functional relationship also provides an explainable relationship between the functional activation and the anatomically defined regions.

Impact Statement

Resting-state functional magnetic resonance imaging labeling using anatomical atlas matching streamlines the process of automatically labeling brain networks through learned spatial representations. By aligning the functional connectivity patterns obtained through independent component analysis (ICA) with anatomical atlases, this method allows for the automatic assignment of labels to the components/networks based on assessing the distribution of the strength of signals across regions. Through this approach, researchers can efficiently identify and interpret the various components extracted via ICA, enhancing our understanding of the brain’s intrinsic connectivity architecture with minimal manual intervention.