Learning Optimal Spectral Clustering for Functional Brain Network Generation and Classification.

Area of Science:

• Neuroscience
• Computational Biology
• Machine Learning

Background:

• Functional brain network (FBN) analysis is crucial for understanding brain organization and diagnosing neurological/psychiatric disorders.
• FBNs possess a small-world topology with functional clusters, where abnormalities are linked to disease.
• Current methods often fail to fully exploit this topology, limiting performance and interpretability.

Purpose of the Study:

• To propose a novel framework, Learning Optimal Spectral Clustering (LOSC), that integrates FBN generation, clustering, and classification.
• To exploit the small-world topology of FBNs through a graph theory-grounded loss function.
• To improve the accuracy and interpretability of FBN analysis for disease diagnosis.

Main Methods:

• LOSC learns brain connectivity in a nonlinear spatio-spectral embedding space using a proposed Rayleigh Quotient Loss (RQL).
• The framework preserves small-world properties in generated FBNs.
• It partitions FBNs into functional clusters and utilizes intra- and inter-cluster relations for classification.

Main Results:

• LOSC achieved consistent accuracy gains of 2.0%, 3.6%, and 2.6% on the ABIDE, ADHD-200, and HCP datasets, respectively.
• The proposed RQL bridges graph theory and learning-based FBN analysis.
• Discovered functional clusters align with known neuropathology and aid in identifying new biomarkers.

Conclusions:

• LOSC offers improved brain network classification accuracy by effectively leveraging small-world functional clusters.
• The framework provides theoretical grounding by integrating graph theory principles into machine learning.
• LOSC enhances interpretability in FBN analysis, aiding in biomarker discovery for neurological and psychiatric disorders.