SELF-CLUSTERING GRAPH TRANSFORMER APPROACH TO MODEL RESTING STATE FUNCTIONAL BRAIN ACTIVITY
Bishal Thapaliya1,2, Esra Akbas1, Ram Sapkota1,2
1Department of Computer Science, Georgia State University, Atlanta, USA.
Summary
A new Self Clustering Graph Transformer (SCGT) method improves brain subnetwork analysis using resting-state fMRI data. SCGT enhances predictions for cognitive scores and gender classification by capturing brain functional connectivity more effectively.
Area of Science:
- Neuroscience
- Machine Learning
- Brain Imaging
Background:
- Resting-state functional magnetic resonance imaging (rs-fMRI) is crucial for understanding brain organization and cognitive processes.
- Graph transformers traditionally use uniform node updates, which may not optimally capture complex brain subnetworks.
Purpose of the Study:
- Introduce a novel attention mechanism, Self Clustering Graph Transformer (SCGT), for graph transformers.
- Address limitations of uniform node updates in graph transformers for brain subnetwork analysis.
Main Methods:
- Developed SCGT, a novel attention mechanism for graphs with subnetworks.
- Utilized static functional connectivity (FC) correlation features as input.
- Applied SCGT to the Adolescent Brain Cognitive Development (ABCD) dataset (7,957 participants).
Main Results:
- SCGT effectively captures and interprets brain subnetwork structures through cluster-specific node updates.
- SCGT outperformed vanilla graph transformers and other recent models in predicting total cognitive score and gender.
- Demonstrated SCGT's efficacy on a large-scale dataset.
Conclusions:
- SCGT offers a promising advancement for modeling brain functional connectivity.
- The method provides enhanced interpretability of underlying subnetwork structures.
- SCGT represents a valuable tool for neuroscience research and brain-related predictions.
