Accurate, markerless optical tracking of head pose for motion correction: A dual-camera simulation with deep learning

Motivation: Deep learning methods for head pose estimation may enable accurate, markerless optical tracking (OT), overcoming practical limitations of OT for motion correction in clinical MRI. Goal(s): To compare the ability of three neural networks to track incremental changes in head pose in 6 degrees of freedom (6DOF) with sub-millimetre/sub-degree accuracy. Approach: We generated a dataset of 20 heads in a simulated MRI environment with in-bore, dual-camera markerless OT and pre-trained the networks prior to training on a real-world dataset. Results: The twin neural network had the lowest test loss (0.13 mm/° across all 6DOF) showing merit in the approach. Impact: Accurate, markerless OT is feasible in simulations with two in-bore cameras and deep learning. Pre-training of a twin neural network was successful (mean RMSE = 0.13 mm/degrees) motivating additional development in the real world, towards motion correction in MRI.