Pairing-based ensemble classifier learning using convolutional brain multiplexes and multi-view brain networks for early dementia diagnosis

Abstract

The majority of works using brain connectomics for dementia diagnosis heavily relied on using structural (diffusion MRI) and functional brain connectivity (functional MRI). However, how early dementia affects the morphology of the cortical surface remains poorly understood. In this paper, we first introduce multi-view morphological brain network architecture which stacks multiple networks, each quantifying a cortical attribute (e.g., thickness). Second, to model the relationship between brain views, we propose a subject-specific convolutional brain multiplex composed of intra-layers (brain views) and inter-layers between them by convolving two consecutive views. By reordering the intra-layers, we generate different multiplexes for each subject. Third, to distinguish demented brains from healthy ones, we propose a pairing-based ensemble classifier learning strategy, which projects each pair of brain multiplex sets onto a low-dimensional space where they are fused, then classified. Our framework achieved the best classification results for the right hemisphere 90.8% and the left hemisphere 89.5%.