Resting state fMRI (rsfMRI) is widely used to differentiate Alzheimer's Disease (AD) and identify biomarkers but its obscure features and noises challenge the present models. Brain graph convolution network (GCN) provides a good interpretation but suffers from the inferior performance due to the insufficient feature representation. Population GCN improves the precision of detection by involving the phenotypic information but fails in the bio logical interpretation. The GCN taking a single generated connectome as input focuses only on the low-level inter regional temporal correlation and is incapable to exploit hierarchical spatial functional features. In this paper, we propose a multi-level connectome-generated GCN (MLC GCN) to enhance the feature extraction for the individual connectome. First, we construct multiple connectomes in parallel through stacked spatiotemporal feature extractors (STFEs), effectively enhancing the hierarchical features and reducing the noise. Each generated connectome is then input into the GCN for further feature extraction, and the output of all GCNs is concatenated for a multilayer percep tron to predict AD. We use independent cohort validations ontwomedicaldatasetsADNIandOASIS-3,andexperiment results demonstrate MLC-GCN obtains better performance for differentiating normal control, mild cognitive impairment and AD than current GCN architectures and other AD classifiers. The proposed MLC-GCNrevealshighinterpreta tion in terms of learning clinically reasonable connectome nodes and connectivity features.
Fu et al. (2026) studied this question.