RSS-Based Machine-Learning-Assisted Localization and Tracking of a Wireless Capsule Endoscope

Abstract

Precise localization of a wireless capsule endoscope (WCE) in the gastrointestinal (GI) tract is paramount for accurate identification of lesions and targeted drug delivery. However, tracking a WCE with high accuracy remains a challenging task. This study presents a WCE localization system with high accuracy and a low root-mean-square error (RMSE) that can localize and track a capsule inside the GI tract with a resolution of 1 cm. The proposed system is based on a comprehensive collection of measured received signal strength (RSS) in a saline-filled American Society for Testing and Materials (ASTMs) phantom. A conformal capsule transmitter, along with an optimized configuration of four on-body receiver antennas operating in the industrial, scientific, and medical (ISM) band at 2.45 GHz, is connected to software-defined radios (SDRs). This setup enables the collection of a substantial dataset comprising 11400 RSS data points, which are systematically mapped to determine the capsule’s position. Data-driven frameworks incorporating three different machine learning (ML) regression models: k-nearest neighbors (KNNs), support vector regression (SVR), and adaptive boosting (AdaBoost), are employed to improve positional accuracy in the localization and tracking processes. Among the utilized ML models, AdaBoost exhibited significant performance with a positional accuracy of 92.60% and an RMSE of 2.38 cm. Moreover, the AdaBoost regression model emerged as the most proficient in tracking a realistic intestinal trajectory with an RMSE of 0.38 cm. Considering its remarkable accuracy, the proposed ML-assisted system is a potential candidate for accurate localization and tracking of a capsule within the GI tract.