Related Experiment Video
Updated: Jun 22, 2025
08:57
Measuring the Influence of Magnetic Vestibular Stimulation on Nystagmus, Self-Motion Perception, and Cognitive Performance in a 7T MRT
Published on: March 3, 2023
1.9K
Test Platform for Developing New Optical Position Tracking Technology towards Improved Head Motion Correction in
Marina Silic1,2, Fred Tam1, Simon J Graham1,2
1Physical Sciences Platform, Sunnybrook Research Institute, Toronto, ON M4N 3M5, Canada.
Sensors (Basel, Switzerland)
|June 27, 2024
Summary
This study introduces a markerless deep learning system for accurate head pose tracking, overcoming limitations of traditional marker-based methods for magnetic resonance imaging motion correction. The new system shows high feasibility for clinical use.
Area of Science:
- Medical Imaging
- Computer Vision
- Machine Learning
Background:
- Optical tracking with fiducial markers aids magnetic resonance imaging (MRI) motion artifact correction but faces clinical adoption barriers due to setup time and calibration.
- Markerless deep learning for head pose estimation has not been applied to MRI motion correction due to stringent sub-millimetre spatial resolution requirements.
- Existing marker-based systems suffer from occlusion, attachment issues, lengthy calibration, and inconsistent performance across degrees of freedom (DOF).
Purpose of the Study:
- To develop and evaluate a markerless, deep learning-based optical tracking system for precise head pose estimation in MRI.
- To overcome the clinical implementation challenges associated with marker-based optical tracking systems.
- To demonstrate the feasibility of a deep learning approach for sub-millimetre accurate head pose tracking.
Main Methods:
- Developed a dual-system approach: a high-fidelity marker-based system for ground truth and a markerless deep learning system.
- Utilized a custom moiré-enhanced fiducial marker for precise ground truth measurements.
- Created a synthetic head pose dataset for initial training of a convolutional neural network (CNN).
Main Results:
- The ground truth system achieved head pose tracking accuracy below 1 mm and 1°.
- Pre-training the CNN on a synthetic dataset yielded an average root-mean-squared error of 0.13 (mm/°) for included models and 0.36 (mm/°) for excluded models.
- The markerless system demonstrated high feasibility for accurate head pose estimation.
Conclusions:
- A markerless deep learning approach is highly feasible for precise head pose tracking in MRI applications.
- This technology has the potential to significantly improve the clinical adoption of optical tracking for motion artifact correction.
- Further research will focus on training and testing the system with real-world data in an MRI environment.
