Through the looking glass: Deep interpretable dynamic directed connectivity in resting fMRI.

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• Functional connectivity (FC) analysis using Pearson correlation is standard for brain network interactions but often uses fixed data windows.
• Deep learning models offer flexible data representation but lack interpretability, requiring post-hoc methods like saliency maps.
• Existing methods struggle with dynamic connectivity estimation and interpretability, hindering clinical application.

Purpose of the Study:

• To develop an interpretable deep learning architecture for brain connectivity analysis.
• To improve accuracy in discriminating between healthy controls and patients with neurological disorders (schizophrenia, autism, dementia).
• To enable accurate prediction of age and gender from functional MRI data.

Main Methods:

• Introduced a deep learning architecture with an interpretable directed graph layer to represent learned brain connectivity.
• Developed a method to estimate windowing functions from data, resolving the fixed window size selection problem for dynamic directed connectivity.
• Compared the proposed model's efficacy against existing classification-focused models using functional MRI data.

Main Results:

• Achieved significantly improved accuracy in discriminating controls and patients with schizophrenia, autism, and dementia, as well as age and gender prediction.
• Demonstrated superior robustness to confounding factors compared to standard dynamic functional connectivity models.
• Identified specific connectivity patterns (e.g., sensorimotor vs. default-mode networks) as key indicators for dementia and gender, and dysconnectivity for schizophrenia.

Conclusions:

• The proposed interpretable deep learning approach enhances brain connectivity analysis for both clinical diagnosis and demographic prediction.
• Task-specific directed connectivity matrices can be estimated, providing naturally interpretable dynamic patterns crucial for prediction.
• The method offers a more robust and insightful alternative to traditional functional connectivity and less interpretable deep learning models.