Matched signal detection on graphs: Theory and application to brain imaging data classification.

Area of Science:

• Graph Signal Processing
• Machine Learning
• Biomedical Data Analysis

Background:

• Recent advancements in signal processing on graphs have opened new avenues for analyzing complex data structures.
• Traditional signal detection methods often fall short when dealing with signals possessing intrinsic graph-based structures.

Purpose of the Study:

• To develop a novel Matched Signal Detection (MSD) theory tailored for signals with graph-based intrinsic structures.
• To extend MSD to handle various signal types, including those in subspaces, with bounded variation, or from prior distributions.
• To evaluate the efficacy of the proposed MSD approach on both simulated and real-world brain imaging datasets for Alzheimer's disease (AD) detection.

Main Methods:

• Graph Laplacian eigenvalues and eigenvectors are utilized to define signal frequencies and subspaces.
• The study explores matched subspace detectors for signals within specific subspaces.
• For more general signals, weighted energy detectors and signal variation differences are employed, incorporating graph Laplacian properties.

Main Results:

• The developed MSD theory effectively processes signals with graph structures.
• Application to Alzheimer's disease (AD) detection using positron emission tomography (PET) and resting-state functional magnetic resonance imaging (R-fMRI) data showed promising results.
• The MSD approach demonstrated superior performance compared to traditional methods in classifying AD and early mild cognitive impairment (MCI) cases.

Conclusions:

• The proposed Matched Signal Detection (MSD) framework offers a powerful tool for analyzing graph-structured signals.
• The MSD approach shows significant potential for early detection of neurodegenerative diseases like Alzheimer's disease by leveraging the manifold structure of brain imaging data.