MAMF-GCN: Multi-scale adaptive multi-channel fusion deep graph convolutional network for predicting mental disorder

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Current mental disorder diagnoses lack objectivity, relying on subjective symptoms and scales.
• Functional magnetic resonance imaging (fMRI) offers potential for objective diagnosis but presents challenges in data integration and analysis.
• Graph neural networks (GNNs) show promise for processing complex relational fMRI data, yet deep integration of multi-modal and multi-scale features remains a challenge.

Purpose of the Study:

• To develop an objective diagnostic method for mental disorders using fMRI data.
• To address the limitations of existing Graph Convolutional Network (GCN) methods in integrating multi-modal and multi-scale features.
• To improve the depth of feature learning and overcome the constraints of single-atlas approaches in GCN models.

Main Methods:

• Proposed a multi-scale adaptive multi-channel fusion deep graph convolutional network (MAMF-GCN) with an attention mechanism.
• Utilized an encoder to combine imaging and non-imaging data, generating similarity weights between subjects.
• Employed multi-channel processing across different brain atlases to extract multi-scale imaging features and fused them using adaptive convolution within a deep GCN.

Main Results:

• The MAMF-GCN method demonstrated superior node classification performance compared to state-of-the-art methods on the Autism Brain Imaging Data Exchange (ABIDE) and Major Depressive Disorder (MDD) datasets.
• Achieved performance improvements of 3.37%-39.83% for MDD and 12.59%-32.92% for ABIDE dataset disease prediction tasks.
• Showcased effective performance in real-life clinical diagnosis scenarios, validating its practical utility.

Conclusions:

• The MAMF-GCN effectively extracts and fuses multi-scale embeddings from imaging features and phenotypic information using an attention mechanism.
• The proposed method significantly enhances classification performance for brain disorder diagnosis.
• This approach offers a robust and objective diagnostic tool for mental disorders by leveraging advanced GNN techniques on fMRI data.