Generalized Recurrent Neural Network accommodating Dynamic Causal Modeling for functional MRI analysis.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Dynamic Causal Modeling (DCM) is a biophysical model linking stimuli to fMRI signals.
• Representing complex stimuli for DCM is challenging.
• Existing Recurrent Neural Networks (RNNs) in fMRI lack interpretability.

Purpose of the Study:

• To develop a biophysically interpretable RNN for fMRI analysis.
• To integrate deep learning with DCM for complex cognitive tasks.
• To enhance the interpretability of RNNs in neuroscience.

Main Methods:

• Proposed a novel DCM-based RNN, termed DCM-RNN.
• Generalized vanilla RNNs to incorporate DCM principles.
• Utilized backpropagation for parameter estimation.
• Demonstrated DCM-RNN's validity in causal architecture and effective connectivity analyses.

Main Results:

• DCM can be represented as a specialized generalized RNN.
• DCM-RNN shows face validity in causal brain architecture and effective connectivity.
• Construct validity was demonstrated in an attention-visual experiment.
• Enabled end-to-end training of DCM and representation learning networks.

Conclusions:

• DCM-RNN offers a promising, interpretable tool for neuroscience research.
• Facilitates seamless integration with classical DCM studies.
• Extends DCM applications to complex tasks like naturalistic stimuli processing.