Computer Vision News - November 2022

53 Roza Güneş Bayrak Developing Computational Models to Enable Individual Precision Modern brain imaging techniques provide non-invasive windows into the anatomy and function of the brain. Since its inception, magnetic resonance imaging (MRI) has become a workhorse for investigating the human brain in health and disease. The primary focus of MRI has been to study similarities across subjects and to make inferences about groups of individuals. However, one movement in brain research is the study of individual precision. In my work, I leverage computational methods to model and interpret large-scale and high dimensional neuroimaging data at the level of individuals. Mechanisms of Functional MRI Signals Brain-body interactions underpin key functions including cognition and emotion, as well as the overall health of an organism. Natural fluctuations, such as breathing and heart rate, provide windows into these critical functions. Yet, many studies of the human brain using fMRI lack these physiological measurements. I developed and tested frameworks [1 ,2 ] to reconstruct respiratory variation (RV) and heart rate (HR) directly from whole-brain fMRI dynamics. To model our physiological signals of interest, I adapted an LSTM network to jointly learn these signals Fig.1 [3] . To achieve generalizability, we utilized large-scale datasets for training; and tested our models on external datasets. The models successfully infer both respiration and heart rate fluctuations; are transferable to scans of variable lengths and different experimental conditions. During the process of investigating our physiological models, we also asked if individuals have unique physiological signal patterns? To assess this, we conducted a series of experiments to relate individual physiological maps to cognitive phenotypes and behavioral traits. We found that physiological variance in fMRI signals correspond to unique properties of individual subjects and are predictive of individual-specific traits. Figure 1: Framework for extracting respiration variation (RV) and heart rate (HR) signals directly from functional MRI data.