Image-based Motion Estimation for Teloperated Flexible Endoscopes

Abstract

Colorectal cancer is one of the leading causes of cancer-related mortality worldwide, although it is almost always preventable by compliance with recommended early screening guidelines. Teleoperated flexible endoscopes have the potential to increase compliance and additionally relieve strain from the physician. Vision-based pose detection systems offer strong advantages for robotic control systems, since they require no additional sensors or augmentation of the endoscope. In this work, we present an algorithm which uses sequential endoscopic images to calculate optical flow (OF), and then train artificial neural networks (ANNs) to estimate metric change in 6 degree of freedom pose from the OF. We investigated the role of white light illumination (WLI) versus narrow band illumination (NBI) and OF representation on the estimation capabilities of the ANNs. Additionally, we analyzed the strength of the features, the distinguishability of the OF patterns, and methods for OF pattern dimensionality reduction. We found that NBI combined with partitioning based on the anatomy of the colon was slightly better for motion estimation in a clinical scenario in which an expert gastroenterologist (>2,000 lifetime procedures) performed colonoscopies on a colon simulator model. We also found that the NBI features were twice as strong as the WLI features. We lastly found that the role of representation of the OF has a significant effect on the estimation capabilities of the ANNs, and that dimensionality reduction does not have a large effect on the estimations, although it reduces the computational load of the algorithm.