A Plug-In Graph Neural Network to Boost Temporal Sensitivity in fMRI Analysis.

Area of Science:

• Neuroimaging
• Machine Learning
• Computational Neuroscience

Background:

• Deep learning models show promise for classifying high-dimensional functional MRI (fMRI) data using functional connectivity (FC) features.
• Existing models often use static FC features, limiting their ability to capture dynamic brain activity patterns.
• Temporal sensitivity is crucial for improving classification accuracy in fMRI analysis.

Purpose of the Study:

• To introduce GraphCorr, a novel graph neural network (GNN) plug-in designed to enhance input features for baseline classification models.
• To improve the performance of classification models by incorporating temporally rich FC features.
• To maintain computational efficiency while leveraging dynamic brain activity information.

Main Methods:

• GraphCorr utilizes a GNN to compute latent FC features with enhanced temporal information.
• It employs a transformer encoder for a node embedder module to capture dynamic BOLD signal representations.
• A lag filter module accounts for delayed interactions by learning correlational features across time delays.

Main Results:

• GraphCorr successfully generates latent FC features with enhanced temporal information.
• The plug-in maintains comparable dimensionality to static FC features.
• Augmenting state-of-the-art graph and convolutional models with GraphCorr led to improved classification performance across three public datasets.

Conclusions:

• GraphCorr offers a novel and efficient method for enhancing fMRI classification by incorporating dynamic FC features.
• The proposed plug-in significantly improves the performance of existing deep learning models.
• GraphCorr represents a valuable advancement for leveraging temporal dynamics in brain activity analysis.