Alzheimer's Imaging Consortium

Area of Science:

• Neuroimaging
• Machine Learning
• Neurology

Background:

• Alzheimer's Disease (AD) is a leading cause of dementia.
• Detecting early AD neurodegeneration via cortical atrophy patterns in T1 MRI is challenging due to subtle, heterogeneous changes.
• Conventional deep learning methods struggle to capture these early, subtle patterns.

Purpose of the Study:

• To develop an explainable cortical graph convolutional network (GCN).
• To capture early atrophy patterns on the cortical surface for identifying subjects at elevated AD risk.
• To improve early diagnosis of prodromal AD.

Main Methods:

• Utilized T1 MRI data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset (1645 subjects).
• Processed MRI data using FreeSurfer and employed a cortical GCN model treating the cortical surface as a graph.
• Trained five-fold cross-validated models on 90% of CN+AD data, testing on 10%, and evaluated prediction of stable MCI (sMCI) vs. progressive MCI (pMCI).

Main Results:

• The cortical GCN model achieved 0.736 balanced accuracy in differentiating dementia (AD) from cognitively normal (CN) subjects.
• The model obtained an average mean balanced accuracy of 0.644 for predicting sMCI from pMCI.
• Demonstrated the model's capability to predict future dementia onset in at-risk individuals.

Conclusions:

• The study demonstrates the effectiveness of cortical GCN for early prediction of dementia onset in individuals at risk for AD.
• Future work includes independent validation on NACC/ADRC data to assess generalizability.
• Enhancing model explainability using techniques like Grad-CAM and integrated gradients is planned to provide deeper insights and improve diagnostic tools.