Predicting high-resolution brain networks using hierarchically embedded and aligned multi-resolution neighborhoods

Abstract

Several works have been dedicated to image super-resolution (i.e., synthesizing high-resolution data from low-resolution data). However, existing works only operate on images (e.g., predicting 7T-like magnetic resonance image (MRI) from 3T MRI) whereas brain connectivity network super-resolution remains unexplored. To fill this gap, we propose the first framework for predicting high-resolution (HR) brain networks from low-dimensional (LR) brain networks by hierarchically aligning and embedding LR neighborhood centered at the testing sample, along with its corresponding HR neighborhood. The proposed hierarchical embedding better preserves higher-order structural neighborhood of subjects within each domain. Recently, a seminal work was introduced for brain network prediction at a single resolution (or scale), where domain alignment was achieved using canonical correlation analysis followed by manifold learning to identify the most similar neighbors to the testing subject (i.e., testing neighborhood) in the source domain that can best predict the missing target network. Here, we inductively extend this idea by hierarchically learning the embedding and alignment of embedding of LR and HR neighborhoods. Our proposed framework achieved the best results in comparison with baseline methods.