Overview

Brain-computer interfaces (BCIs) create a direct communication channel between the nervous system and external devices. Our lab focuses on a particularly challenging and impactful frontier: using visual imagery as the information carrier in BCI systems. By decoding the neural correlates of what a user sees or imagines, we can build richer, more intuitive interfaces than those based on simple motor commands.

Research Directions

Visual Image Reconstruction from Neural Signals We apply latent diffusion models conditioned on EEG/fMRI embeddings to reconstruct the images a subject perceives. Our two-stage pipeline first maps neural signals to a CLIP-compatible semantic embedding space, then uses a fine-tuned Stable Diffusion decoder to synthesize the image. On the THINGS-EEG2 benchmark, our method achieves a top-1 image retrieval accuracy of 43.7% (chance: 0.5%).

Steady-State Visually Evoked Potential (SSVEP) Spellers For patients with amyotrophic lateral sclerosis (ALS) and other locked-in conditions, we develop high-throughput SSVEP-based communication boards. We combine frequency-domain feature extraction with a lightweight CNN classifier, achieving an information transfer rate of 68 bits/min with a 40-class stimulus matrix—competitive with the state of the art while requiring only consumer-grade EEG hardware.

Closed-Loop Neurofeedback for Attention Restoration Leveraging real-time alpha/theta power estimates, our neurofeedback system generates personalized visual stimulation schedules to sustain user attention. A randomized pilot study (n=24) showed a statistically significant 18% improvement in sustained attention task accuracy after four sessions.

Current Status

  • Image reconstruction pipeline: published, code released
  • SSVEP speller: system prototype complete; clinical usability study underway
  • Neurofeedback: pilot data collected; full trial IRB approved, enrollment ongoing

Key Collaborators

  • State Key Laboratory of Bioelectronics
  • Department of Rehabilitation Medicine

Selected Output

  • Journal paper in IEEE Transactions on Neural Systems and Rehabilitation Engineering (2024)
  • Best Paper Award, IEEE EMBC 2023
  • Open-source toolbox: NIM-BCI (PyTorch-based EEG decoding framework)