Can functional connectivity be used to refine structural connectivity strength by combining neural computational model and generative adversarial network?

Currently, structural connectivity (SC) strength is usually non-invasively estimated as streamline density or connection probability, via tractography based on diffusion magnetic resonance imaging (dMRI). However, nonnegligible tracking biases are still unavoidably introduced into SC strength computation, thus affecting exploration of anatomical mechanism underlying functional interactions. Here, we refined SC strength from empirical functional connectivity (FC) by combining neural computational model and deep-learning techniques. First, the neural computation model (Generic2dOscillator) was employed to generate simulated FCs from dMRI-tracked SCs. Then, these two kinds of connectivity matrices were used as samples to train generative adversarial network (GAN) incorporating graph attention (GAT). Last, empirical FCs were fed to the trained GAN to infer SCs. Global topological metrics of the predicted SCs have better consistency, no matter whether brains were parcellated into 62 or 132 subregions. More reasonable SC strength was obtained for the specific structural connections that are prone to tracking bias. These findings help us in understanding of the brain’s white matter organization, and in integrating dMRI and fMRI images to study coupling relationship between SC and FC networks.