Research

Exploring the intersection of deep learning and neuroimaging

Research Interests

Deep Learning for Neuroimaging

Developing and applying deep learning models for analyzing various neuroimaging modalities including fMRI, structural MRI, and diffusion MRI.

Structural Connectivity Analysis

Investigating white matter pathways and structural connectivity patterns in the brain, particularly in visual processing regions.

fMRI Transformer Models

Building foundation models for functional MRI data analysis using transformer architectures, enabling end-to-end analysis of brain dynamics.

Psychiatric Disorder Diagnosis

Using machine learning and neuroimaging data to diagnose and predict trajectories of psychiatric disorders, identifying biomarkers for symptoms.

Generative Models for Neuroimaging

Applying generative adversarial networks and other deep learning models to correct artifacts, synthesize data, and enhance neuroimaging analysis.

Research Projects

Deep Learning Tract Segmentation Framework

Pestilli Lab, University of Texas at Austin • Aug 2024 – Present

Implemented neural tract segmentation using foundation models, benchmarking performance against established methodologies.

  • •Implemented neural tract segmentation using foundation models, benchmarking performance against established methodologies to demonstrate comparative advantages.
  • •Built a comprehensive analysis pipeline on the Brainlife platform, enabling efficient model training and validation across multiple diffusion MRI datasets.

VISCONTI: Structural Connectivity Analysis Pipeline for Early Visual Cortex

Pestilli Lab, University of Texas at Austin • Aug 2024 – Present

Developed pipeline integrating tractography with population receptive field mapping to quantify structural connectivity patterns.

  • •Developed the VISCONTI pipeline, integrating tractography with population receptive field mapping to quantify structural connectivity patterns in the early visual cortex.
  • •Revealed structural connectivity basis for visual performance asymmetry through systematically analyzing white matter pathways connecting visual processing regions.

SwiFT: Swin 4D fMRI Transformer

Connectome Lab, Seoul National University • Dec 2020 – Jul 2024

Developed a deep learning model for end-to-end fMRI data analysis, surpassing existing methods in predicting biological and cognitive variables.

  • •Developed a deep learning model (SwiFT) for end-to-end fMRI data analysis, surpassing existing methods in predicting biological and cognitive variables.
  • •Used SwiFT to predict task-related brain activity from resting-state data, identifying strong correlations with personal attributes like neuroticism and depressive symptoms.

Diagnosing and Predicting Future Trajectories of Psychiatric Disorders

Connectome Lab, Seoul National University • Dec 2020 – Jul 2024

Employed clustering algorithms and machine learning models to identify EEG biomarkers for psychiatric symptoms.

  • •Employed clustering algorithms to categorize psychiatric symptoms and used machine learning models to identify corresponding Electroencephalography (EEG) biomarkers.
  • •Preprocessed EEG data and integrated multi-modal neuroimaging data with machine learning in the EMBARC study.

Enhancing Neuroimaging with Generative Deep Learning Models

Connectome Lab, Seoul National University • Dec 2020 – Jul 2024

Applied generative models for MRI data correction and synthesis, including GANs and Fourier Neural Approximators.

  • •Corrected site effects in MRI data with cycle-consistent Generative Adversarial Networks (GANs).
  • •Generated diffusion MRI data from existing structural data using frequency-aware GANs.
  • •Introduced Fourier Neural Approximator for synthesizing bandlimited signals like EEG.

Research Summary

My research focuses on developing and applying advanced machine learning and deep learning techniques to neuroimaging data. I work with various neuroimaging modalities including functional MRI (fMRI), structural MRI, and diffusion MRI to understand brain structure and function.

A significant portion of my work involves developing foundation models for neuroimaging data analysis. My work on SwiFT (Swin 4D fMRI Transformer) demonstrates how transformer architectures can be adapted for spatiotemporal brain data, enabling end-to-end analysis that surpasses traditional methods in predicting biological and cognitive variables.

I am also interested in structural connectivity analysis, particularly in understanding how white matter pathways connect different brain regions. My recent work on the VISCONTI pipeline integrates tractography with population receptive field mapping to reveal structural connectivity patterns in the early visual cortex.

Additionally, I apply machine learning techniques to psychiatric disorder diagnosis and prediction, using multi-modal neuroimaging data to identify biomarkers and predict disease trajectories. I also work on generative models for neuroimaging, using GANs and other deep learning approaches to correct artifacts, synthesize data, and enhance analysis pipelines.