Network analysis on neuro-imaging data
This study explored the design of a complete, explainable, scalable, and reliable brain network construction pipeline using Diffusion Tensor Imaging (DTI) images to generate structural connectivity matrices for analysis; this analysis included the evaluation of various Graph Convolutional Network (G...
Saved in:
Main Author: | |
---|---|
Other Authors: | |
Format: | Final Year Project |
Language: | English |
Published: |
Nanyang Technological University
2024
|
Subjects: | |
Online Access: | https://hdl.handle.net/10356/175076 |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Institution: | Nanyang Technological University |
Language: | English |
Summary: | This study explored the design of a complete, explainable, scalable, and reliable brain network construction pipeline using Diffusion Tensor Imaging (DTI) images to generate structural connectivity matrices for analysis; this analysis included the evaluation of various Graph Convolutional Network (GCN) baselines including various hierarchical pooling (DiffPool, SAGPool, ASAP) and readout (SoftmaxAggregation, PowerMeanAggregation) modules. A total of 79 subjects from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) from two classes of Alzheimer’s disease classification, Cognitive Normal (CN) and Late Mild Cognitive Impairment (LMCI), were included in this study. Overall, it was found that the SCHAEFER100 (Schaefer 7 networks 100 parcellations) atlas fared better than AAL116 (AAL 116 parcellations) atlas across all experiments and that hierarchical pooling has a positive impact on graph classification in the case of Schaefer. Surprisingly, the global sum aggregation readout was sufficient to capture the node
feature information of the brain graphs, with the best-performing learnable readout function SoftmaxAggregation faring slightly below it in terms of test accuracy. The impressive model performance across all baselines also proves the high quality of the structural connectivity matrices generated. This study is significant in helping researchers new to computational neuroscience to quickly grasp the important parts of brain network construction and serves as a preliminary exploration of how graph pooling and readout modules are crucial in model design for brain network analysis. |
---|