Machine Learning Assisted Non-Rigid Surface Tracking in Biological Systems

Abstract

Occlusions, obstructions, and lighting changes that occur in a camera's field-of-view (FOV) during a medical procedure can cause tracking algorithms to lose track of a particular region-of-interest (ROI). Various approaches to reacquire the tracking of rigid objects have been developed, however, the non-rigid nature of biological structures requires more complicated approaches. The purpose of this research is to improve the performance of an existing non-rigid tracking algorithm under these types of adverse conditions. This existing algorithm was previously shown to be accurate and efficient under ideal conditions but exhibited a high rate of tracking failures due to the aforementioned occlusions, obstructions, and lighting changes. To improve tracking under these conditions, a tissue motion machine learning model was developed to provide predictions of future ROI grid motion. The combination of this machine learning technique along with various improvements to the base algorithm was shown to greatly reduce the number of tracking resets and allow the tracking grid to briefly follow an expected motion pattern during a simulated occlusion.