Group-representative functional network estimation from multi-subject fMRI data via MRF-based image segmentation.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Functional connectivity patterns from fMRI data are increasingly used to characterize group traits.
• Identifying consistent group patterns across subjects is challenging due to existing methods' limitations.

Purpose of the Study:

• To develop a principled and flexible approach for accurate identification of group-representative functional connectivity patterns.
• To overcome limitations of existing methods in fMRI data analysis.

Main Methods:

• Redefines functional network estimation as an image segmentation problem.
• Utilizes a maximum a posteriori-Markov random field (MAP-MRF) framework with a probabilistic model.
• Employs a novel variational Bayes algorithm to recover latent group images from fMRI data.

Main Results:

• The proposed model with variational Bayes outperforms other algorithms, even with model misspecification.
• Demonstrates robustness to initialization and ability to recover ground truth in simulated fMRI data.
• Successfully identifies subtle group differences in fMRI data from control and depression groups.

Conclusions:

• Probabilistic models and algorithms accounting for uncertainty are advantageous for identifying group connectivity maps.
• Variational Bayes offers accurate results with no increased computational load and robustness to model misspecification.
• The methodology improves the ability to avoid local optima in solutions.