Regression DCM for fMRI

Area of Science:

• Computational Neuroscience
• Neuroimaging Analysis
• Network Science

Background:

• Inferring effective connectivity in large-scale neural networks from neuroimaging data is a significant computational challenge.
• Traditional Dynamic Causal Models (DCMs) are limited to small networks (approx. 10 regions) due to computational constraints during model inversion.
• Existing methods struggle to scale for whole-brain analyses.

Purpose of the Study:

• To develop a novel variant of DCM suitable for assessing effective connectivity in large, whole-brain networks using fMRI data.
• To significantly improve the computational efficiency of inferring effective connectivity in complex neural systems.
• To enable whole-brain connectomics by analyzing large-scale neural interactions.

Main Methods:

• Developed regression DCM (rDCM) by translating linear DCM into the frequency domain, reformulating it as Bayesian linear regression.
• Implemented a variational Bayesian inversion method for extremely fast inference.
• Validated rDCM using simulated and empirical fMRI data across various signal-to-noise ratios (SNR) and repetition times (TR).

Main Results:

• rDCM demonstrated significantly accelerated model inversion, orders of magnitude faster than classical DCM.
• The method showed face validity across different SNR and TR settings in fMRI data.
• Successfully inferred effective connection strengths in a simulated 66-region whole-brain network with 300 parameters.

Conclusions:

• rDCM is a computationally efficient approach for inferring effective connectivity in large-scale neural networks.
• The method shows significant promise for advancing whole-brain connectomics using individual fMRI data.
• rDCM overcomes limitations of traditional DCM, enabling more scalable network analyses.