Prediction of amyloid positivity based on white matter structural integrity in non-demented individuals using 3D CNN

Motivation: There exists a nonmonotonic relationship between water molecule diffusion in white matter and cerebral amyloid deposition. Goal(s): The primary goal of this study is to use deep learning to simultaneously learn spatial and temporal features of white matter microstructure for prediction of amyloid positivity. Approach: (2+1)D convolutional neural networks (CNNs) with different pooling layers are used to learn independent spatiotemporal features from various aspects of diffusion (free water fraction, neurite density index and orientation dispersion). Collective information from these features are then utilized for amyloid positivity prediction. Results: White matter microstructure predicts amyloid burden with a mean validation AUC of 0.7. Impact: Early detection of elevated amyloid burden in non-demented patients using diffusion MRI can stratify patients with a high risk of developing Alzheimer's disease. It could potentially serve a prognostic biomarker for recommending early clinical intervention.