Predicting subject-specific brain functional connectivity from structural connectivity: a deep learning perspective

Abstract

Previous neuroscientific studies have reported strong relationships between brain structure and function. In humans, this has most commonly been done using non-invasive, MRI-based measures of resting-state functional connectivity (FC) and structural connectivity (SC) in combination with computational modeling, and more recently artificial intelligence (AI) methods, to explore SC-FC correlations and attempts to predict one form of connectivity (e.g., FC) from the other (e.g., SC). However traditional computational modeling-based studies evaluated the results based on “conventional parameters” (statistical measurements such as: mean squared error, correlation, variance) which are prone towards floor effect. Due to the tremendous preprocessing steps involved in data processing pipelines, there lies a definitive need for a standard dataset for comparing connectivity-based analysis and evaluating the efficacy of the experimental methods. To overcome these problems, a curated open dataset (containing SC, and different forms of FC), a set of traditional performance parameters along with a novel parameter are proposed. Using a practical example (structural (diffusion MRI) and functional (resting-state fMRI) connectomes from 762 participants in the Human Connectome Project (HCP, S900 release) – regions of interests parcellated from Glasser atlas), we wanted to establish the baseline performance through a Graph Convolutional Network (GCN) \& U-Net. For optimizing the baseline performance and improving the performance of the algorithm, we systematically modified these two different AI-based methods (6 GCN’s and 2 U-Net’s; in total 8 configurations) to establish a more optimized deep learning-based approach for predicting subject-specific brain FC from SC. These eight configurations were evaluated using both conventional performance metrics as well as the recently proposed pairwise functional connectome fingerprinting approach (PFCF) to determine how closely matched each subjects predicted and measured FCs are, relative to inter-subject differences in measured FCs. These results show that for SC-FC prediction, all of the GCN architectures worked better than the U-Nets. Although traditional performance metrics have indicated differences in performance for each structural variation, PFCF is quantitatively more rigorous. Therefore, we conclude that the current analysis based on a comprehensive set of metrics on the standardized dataset will provide a new benchmark for future AI-based connectivity prediction.