A Machine Learning Enabled Near Infrared Tracking Scheme for Localization of Gastrointestinal Smart Capsule

Abstract

Localizing a smart capsule within the gastrointestinal (GI) tract is essential for high performance, accurate sensing as well as efficacious drug delivery at designated locations. In this work, we describe a data-driven framework that employs a near infrared (NIR) tracking scheme to achieve the localization of smart capsule in GI tract. A prototype of the tracking system consists of a single NIR LED of 940 nm incorporated with an array of readout device that integrated with an array of NIR photodiodes. A data-driven approach was applied to build the non-linear estimation model and estimate the capsule’s localization by interpolating the outputs of the photodiodes in response to the movement of the NIR LED. Three different machine learning models: support vector regression (SVR), KNN, adaptive boosting (AdaBoost), and the trilateration positioning approach were trained on calibration data. These models were validated in ex vivo experiments with different thick of porcine tissue (e.g. 20 mm thick) using a NIR LED (three types of intensities) and three photodiodes in different patterns. The capsule localization predicted by these machine learning models showed best estimation results ( <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\text{R}^{2} =96.00$ </tex-math></inline-formula> %, and RMSE = 5.13 mm) when using Adaboost; a second best performance was achieved by SVR with a tradeoff on accuracy and time saving. These results suggest that the proposed machine learning data-driven enabled NIR tracking system can be an effective tool for measuring real-time location of gastrointestinal capsule.