AI/ML in Neuroimaging
We developed and utilized state-of-the-art artificial intelligence (AI) and machine learning (ML) models to study the brain. We decoded visual stimulation and semantic information from brain signals, including fMRI and large-scale neuronal recordings, using AI/ML models. We developed a deep learning model that incorporates a contrastive language–image pre-training (CLIP) encoder to decode the semantic information of image stimuli based on fMRI responses they evoked. We also introduced Brain Masked Auto-Encoder (BrainMAE) for learning representations directly from dynamic fMRI time series. Additionally, we worked with our collaborators to apply multi-graph-based deep networks to neuroimaging research.
Infra-slow global brain activity and its functional relevance
We have demonstrated the existence of a highly structured infra-slow global brain activity. This activity, observed in human fMRI and monkey ECoG, takes the form of infra-slow waves that propagate along the cortical hierarchy gradient, resembling the cross-layer information flow in artificial neuronal networks. In large-scale neuronal recordings from mice, this activity is manifested as spiking cascades of sequential activations across ~70% recorded neurons, regardless of brain regions. This infra-slow global brain activity is phase-coupled with arousal modulations of the seconds scale, as well as the hippocampal ripples, which are known to be critical for memory functions. We have also showed that this cascade dynamic is closely linked to the hippocampal replays of movie-induced sequence in the hippocampus. In human fMRI, this global brain activity, manifested as peaks of global mean signal (gBOLD), is coupled by cerebrospinal fluid (CSF) movement. Their coupling strength is closely associated with various pathologies of Alzheimer’s disease (AD) and cognitive decline in Parkinson’s diseases (PD). Moreover, its propagation is related to the spreading of Amyloid-beta in the early stages of AD. Overall, this global brain activity/dynamic appears to be associated with memory, cholinergic function, and waste clearance processes that are all impaired in AD.
Neural Basis of Resting-state fMRI Connectivity
The correlation of low-frequency (< 0.1 Hz) functional Magnetic Resonance Imaging (fMRI) signals acquired during the resting-state has been widely used for measuring the brain connectivity in healthy and diseased brains. However, the neural basis underlying this measure remains elusive, which makes many significant findings from resting-state fMRI studies remain empirical observations. Our group combined multiple neuroimaging modalities, including fMRI, electroencephalogram (EEG), concurrent measurement of both, and large-scale electrocorticography (ECoG), to investigate the neural basis of resting-state fMRI correlations in humans, monkeys, and rats.
Methods for Quantifying Network Dynamics of the Brain
The “functional” MRI data typically has 4 dimensions with one in time. Conventional analyzing approaches focus on statistics averaged over time and thus ignore dynamic information of brain. Our group develops methods for characterizing instantaneous, repertorial co-activation patterns of the brain and metrics for quantifying network dynamics of the brain, which may be potentially used for differentiating brains of healthy and diseased, under different states, and with different behaviors.
Brain Activity under Different States of Consciousness
Despite the rapid development of modern neuroscience in recent years, we made very small progress in understanding a centuries-old question: what is the neural correlate of consciousness. Our group is trying to tacking this question by charactering and comparing large-scale network activity and connectivity under different states of consciousness (e.g., wake, sleep, and anesthesia), as well as their temporal dynamics at transitions across these states, using modern neuroimaging techniques.