Fyzeen Ahmad

There exist many data-driven methods by which we can use resting-state functional MRI (rsfMRI) data to identify regions of the brain that activate together. These analysis methods produce representations of what is known as the “human connectome,” all connections in the brain that give rise to complex behavior and cognition. Meaningful insights, in both clinical and basic neuroimaging research, have been derived from various representations of the human connectome. However, it is unknown whether we gain fundamentally different insights when using different representations or simply see manifestations of the same underlying neural mechanisms. We attempted to bridge this gap by training geometric deep learning algorithms to translate between different representations of functional connectivity among a single cohort of subjects in the Human Connectome Project (n=1003). Our trained models showed excellent accuracy on translations between ICA and PROFUMO representations of the human connectome and good but limited accuracy on translations to Gradient representations. These results indicate ICA, PROFUMO, and Gradient representations capture many but not all of the same aspects of the human connectome, despite being created with vastly different computational manipulations of the raw rsfMRI data. These findings warrant further investigation into how information about the human connectome stored in these representations actually differ from each other.