Overview

Accurate navigation is critical in neurosurgery, where errors of even one millimeter can lead to irreversible neurological damage. Traditional neuronavigation systems rely on rigid registration between preoperative scans and the intraoperative workspace—an approach that struggles to account for brain shift caused by craniotomy, cerebrospinal fluid drainage, and tissue resection.

Our group develops intelligent neural navigation systems that continuously adapt to intraoperative changes using deep learning and real-time multimodal sensing.

Research Directions

Real-Time Brain Shift Compensation We train deformable registration networks on paired intraoperative ultrasound and preoperative MRI volumes. Our current model achieves mean target registration error (TRE) of 0.71 mm on a held-out clinical dataset, outperforming conventional biomechanical simulation baselines.

Instrument Tracking and Pose Estimation Using stereo endoscopy and electromagnetic tracking fusion, we reconstruct the 3D pose of surgical instruments at >30 Hz. A Transformer-based tracking module fuses both modalities to handle occlusion and line-of-sight failures common in deep surgical corridors.

Surgical Path Planning with Reinforcement Learning We formulate the trajectory optimization problem as a Markov Decision Process (MDP) and train policies in a patient-specific simulation environment derived from segmented CT/MRI. The learned policies avoid critical vascular and eloquent cortex structures while minimizing insertion depth.

Current Status

  • Brain shift compensation model: under clinical validation at partner hospital
  • Instrument tracking pipeline: integrated into lab prototype system
  • Path planning: simulation phase; preclinical animal trials planned for Q4 2025

Key Collaborators

  • Department of Neurosurgery, affiliated teaching hospital
  • School of Biomedical Engineering

Selected Output

  • Conference paper at MICCAI 2024 (brain shift compensation)
  • Patent application filed: intelligent navigation system with intraoperative adaptation
  • Open-source dataset: 120 paired MRI–ultrasound cases (release planned 2025)