Functional Brain Network Classification Based on Deep Graph Hashing Learning.

Area of Science:

• Neuroscience
• Medical Imaging
• Machine Learning

Background:

• Resting-state functional magnetic resonance imaging (rs-fMRI) is crucial for diagnosing brain diseases.
• Deep learning methods for brain network classification are gaining traction.
• Current deep learning approaches often overlook semantic proximity relationships in brain networks.

Purpose of the Study:

• To propose a novel brain network classification method, BNC-DGHL, based on deep graph hashing learning.
• To incorporate semantic spatial relationships into brain network classification.
• To enhance the accuracy of brain disease diagnosis using rs-fMRI data.

Main Methods:

• Extraction of deep features from brain networks.
• Learning a graph hash function utilizing clinical phenotype and diagnostic label similarity.
• Conversion of deep features into hash codes that preserve semantic spatial relationships.
• Classification of brain networks by calculating distances between hash codes.

Main Results:

• The BNC-DGHL method demonstrated superior classification performance on ABIDE I, ABIDE II, and ADHD-200 datasets.
• The identified abnormal functional connectivities may act as potential biomarkers for brain diseases.
• The method effectively maintains original semantic spatial relationships in hash codes.

Conclusions:

• The proposed deep graph hashing learning method offers improved brain network classification for disease diagnosis.
• BNC-DGHL effectively integrates spatial topology and semantic proximity for enhanced diagnostic accuracy.
• The findings suggest potential neuroimaging biomarkers for specific brain diseases.