Geostatistical analysis in clustering fMRI time series

Area of Science:

• Neuroimaging
• Brain Activity Analysis
• Geostatistics

Background:

• Functional magnetic resonance imaging (fMRI) time series clustering is common for brain activity identification.
• Existing methods like direct time series clustering or cross-correlation have limitations.
• Direct clustering may yield unrelated temporal behaviors, while cross-correlation needs protocol knowledge.

Purpose of the Study:

• To propose and validate autocorrelation structure as a superior feature for fMRI time series clustering.
• To formalize traditional classification methods into feature extraction, metric choice, and algorithm choice.
• To investigate the impact of pre-clustering masking on discovered brain activity clusters.

Main Methods:

• Formalized traditional fMRI clustering into feature extraction, metric, and algorithm selection.
• Employed autocorrelation structure, inspired by geostatistics, as a novel feature for clustering.
• Applied autocorrelation clustering to fMRI data from a visual task and resting-state conditions.

Main Results:

• Autocorrelation clustering effectively identifies active brain regions without needing stimulus protocol information.
• Demonstrated the efficacy of autocorrelation clustering on both task-based and resting-state fMRI data.
• Found that pre-clustering masking can potentially degrade the quality of identified brain activity clusters.

Conclusions:

• Autocorrelation clustering offers a robust alternative to traditional methods for fMRI brain activity identification.
• The proposed method overcomes the drawbacks of time series clustering and cross-correlation.
• Pre-clustering masking should be carefully considered as it may negatively impact results.