Alzheimer’s disease classification based on brain region-to-sample graph convolutional network

• Published in: Biomedical signal processing and control 2024-10, Vol.96, p.106589, Article 106589
• Main Authors: Yang, Zhi, Liu, Wenya, Gan, Haitao, Huang, Zhongwei, Zhou, Ran, Shi, Ming
• Format: Article
• Language: English
• Subjects: Alzheimer’s disease; Brain connectivity graphs; Graph Convolutional Network; Magnetic resonance imaging
• ISSN: 1746-8094
• DOI: 10.1016/j.bspc.2024.106589

Alzheimer’s disease (AD) is a notable high prevalence neurodegenerative disorder worldwide. Graph convolutional network (GCN) have emerged as a prominent technique for the classification of AD by leveraging graph-based learning power. Traditional GCN based AD classification methods typically rely on the direct utilization of feature vectors from samples, making the accuracy of predictions heavily dependent on the quality of the input graphs. To address these challenges, we propose a novel region-to-sample graph convolutional neural network framework. This framework firstly constructs a brain connectivity graph directly utilizing MRI data. Subsequently, it employs a RegionGCN module that integrates GCN with node weight adjustment techniques to generate refined feature vectors for each sample. Finally, the SampleGCN module, which merges GCN with cross-attention mechanisms, aggregates neighborhood information from the sample graph, achieving precise classification of samples. Our method has undergone rigorous experimental validation using the ADNI dataset, a dataset in the public domain, and the results demonstrate robust competitiveness in multiple classification tasks related to AD. This confirms that the approach we propose is not only innovative in design but also holds substantial potential for broader applications in the medical field. •RSGCN, consisting of RegionGCN and SampleGCN, enhances the AD classification.•RegionGCN adapts node weights for accurate sample feature extraction.•SampleGCN enhances disease prediction accuracy by integrating multiscale features.