Predicting gender from structural and functional connectome via multi-view graph neural networks

• Main Author: He, Yinan
• Other Authors: Jagath C Rajapakse
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/166152
• Institution: Nanyang Technological University

Neuroimaging data are high dimensional and scarce. This causes overfitting to be a significant problem during model training, which is detrimental to the generalizability of the trained model. Using multiple modalities can introduce valuable complementary information between different neuroimaging methods, but this further aggravates the overfitting problem. Various existing methods used graph neural network (GNN) based models for predictions on graph representations of neuroimaging data and have achieved remarkable results. However, current research has not thoroughly explored the use of multiple modalities for methods involving both brain and population graphs. We propose the BrainGAT, a two-stage early fusion Graph Attention Network method (GAT) for classification that involves the use of both brain and population graphs. Our method uses multiple modalities for subject representation learning and then constructs a population graph to learn inter-subject information. We aim to fully exploit the information from multiple modalities and the subject relationship for prediction. We benchmark the proposed method on gender classification task and demonstrate that our proposed model outperforms all baseline models, achieving an accuracy of 76.76% on Parkinson’s disease dataset PPMI and 82.50% on a larger HCP dataset of healthy young adults.