Residual-ChebNet: A residual connection graph convolutional network for mild cognitive impairment classification based on multimodal imaging data

Mild cognitive impairment (MCI) is characterized by a decline in cognitive ability, but the degree is milder than Alzheimer’s disease (AD). Approximately 15% of MCI patients may progress to AD within three years. Therefore, early diagnosis and intervention of MCI are crucial to delay the progression of AD. However, traditional single-modality computer-aided diagnosis (CAD) methods, such as those based on structural magnetic resonance imaging (sMRI), are limited by the subtlety of MCI symptoms and individual differences, resulting in low diagnostic accuracy and a lack of interpretability in conclusions. To address these issues, this study proposes a series of innovative approaches: (1) A novel residual graph convolutional network (Residual-ChebNet) is proposed. The network architecture, which integrates individual characteristics and population data associations, is designed to enhance the recognition of intrinsic relationships in the MCI population graph. It also address the gradient vanishing problem of traditional convolutional neural networks to optimize MCI diagnostic performance. (2) A multimodal data fusion and graph structure representation method is constructed. For the preprocessed resting-state functional magnetic resonance imaging (rs-fMRI) and diffusion tensor imaging (DTI) data, a population sparse graph is constructed. Through the weighted clustering coefficient, the functional and structural information of the brain connection network is integrated and converted into a graph node feature vector. In summary, our model has the following advantages: (1) A multimodal fusion strategy that combines rs-fMRI functional connectivity with DTI structural connectivity to overcome the representation limitations of single-modal data. (2) A population sparse graph model that captures the common pathological characteristics of the MCI population by modeling population-level neuroimaging associations through graph structures. (3) An interpretability enhancement design that visualizes brain region structural connection information via graph node features, providing a neuroscientific basis for diagnostic conclusions. (4) Experimental validation and performance improvement: Experiments on the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset show that the accuracy of early MCI (EMCI) recognition reaches 0.9286, which is 2.15% higher than that of the state-of-the-art method; the sensitivity reaches 0.9350, which is 2.86% higher. Additionally, the visualization analysis of the model further verifies the effective depiction of brain region connection patterns by graph representation, highlighting the advantages of Residual-ChebNet in the analysis of MCI pathological mechanisms. This study achieves dual improvement in MCI diagnostic accuracy and interpretability through the innovative architecture of multimodal graph convolutional networks, providing a new technological paradigm for the early diagnosis of AD.