Pre-Training Graph Attention Convolution for Brain Structural Imaging Biomarker Analysis and Its Application to Alzheimer’s Disease Pathology Identification

Biomarkers are one of the primary diagnostic tools to facilitate the early detection of Alzheimer’s disease (AD). The accumulation of beta-amyloid (Aβ) plaques in the human brain is one of the presymptomatic hallmarks of AD. However, current methods to detect Aβ pathology are either invasive (lumbar puncture), quite costly, and not widely available (amyloid positron emission tomography - PET), or largely under development (blood-based biomarkers - BBBM). Thus a less invasive and cost-effective approach is demanded. Magnetic resonance imaging (MRI), which has been used widely in preclinical AD, has recently shown the capability to predict brain Aβ positivity. This motivates us to develop a method called pre-training graph attention convolution, which uses MRI data to predict Aβ positivity. Our proposed self-supervised learning architecture refines feature extraction from mesh representation through pre-training and fine-tuning, ultimately yielding more powerful biomarkers for identifying Aβ. We obtain subjects from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) and use our method to discriminate Aβ positivity. Theoretically, we provide analysis toward the understanding of what the network has learned. Empirically, it shows strong performance on par or even better than state of the art.