Predicting brain amyloid-β PET phenotypes with graph convolutional networks based on functional MRI and multi-level functional connectivity

Abstract The detection of amyloid-β (Aβ) deposition in the brain provides crucial evidence in the clinical diagnosis of Alzheimer’s disease (AD). However, the efficiency of the current PET-based brain Aβ examination suffers from both coarse, visual inspection-based bi-class stratification and high scanning cost and risks. In this work, we explored the feasibility of using non-invasive functional magnetic resonance imaging (fMRI) to predict Aβ-PET phenotypes in the AD continuum with graph learning on brain networks. First, three whole-brain Aβ-PET phenotypes were identified through clustering and their association with clinical phenotypes were investigated. Second, both conventional and high-order functional connectivity (FC) networks were constructed using resting-state fMRI and the network topological architectures were learned with graph convolutional networks (GCNs) to predict such Aβ-PET phenotypes. The experiment of Aβ-PET phenotype prediction on 258 samples from the AD continuum showed that our algorithm achieved a high fMRI-to-PET prediction accuracy (78.8%). The results demonstrated the existence of distinguishable brain Aβ deposition phenotypes in the AD continuum and the feasibility of using artificial intelligence and non-invasive brain imaging technique to approximate PET-based evaluations. It can be a promising technique for high-throughput screening of AD with less costs and restrictions.