• ABIDE
• AI
• ASD
• GNN
• Machine Learning
• MRI
• multimodal
• neuroimaging data analysis
• rs-fMRI

This master’s thesis explores the use of Graph Neural Networks (GNNs) for analyzing the human connectome, with a focus on identifying Autism Spectrum Disorder (ASD) using neuroimaging data. The brain is modeled as a graph, with nodes representing anatomical regions and edges representing functional connections. The study develops a pipeline to evaluate GNNs’ predictive capabilities and identify relevant brain subregions for classification. The methodology consists of three stages: feature extraction from MRI data, graph construction to model the brain connectome, and GNN-based analysis for classification and interpretability.

The research utilizes the ABIDE dataset, focusing on male subjects aged 5–35, and extracts morphological features from structural MRI and time-series data from resting-state fMRI. Graph-structured representations are constructed using Pearson Correlation Matrix (PCM). The study frames the task as a binary classification to distinguish ASD from typically developing (TD) individuals. The GNN model, a GraphSAGE-inspired approach (SAGPooling-GraphSAGE), achieves an AUC score of 72.2 ± 1.8, demonstrating its effectiveness in classification compared to traditional Machine Learning models like Random Forest.

Self-attention scores help identify the most relevant brain regions for classification, offering both local and global interpretability. In conclusion, the thesis shows that GNNs can classify ASD vs. TD with high performance and provide valuable insights into the brain’s functional connectome.