Residual Neural Network for Precise 6-DoF Capsule Endoscope Localization Using Electromagnetic Induction

Accurate localization of capsule endoscopes is essential for advancing minimally invasive gastrointestinal diagnostics. Traditional methods based on magnetic induction often struggle with convergence issues and sensitivity to initial conditions, limiting their accuracy and robustness. This paper introduces a deep learning framework that overcomes these limitations by directly estimating the six degrees-of-freedom pose—position (X, Y, Z) and orientation (roll, pitch, yaw)—of a capsule endoscope. Our approach maps nine electromotive force (EMF) measurements, generated by a three-coil transmitter and a tri-axis receiver, to the capsule's precise pose. The proposed model integrates residual blocks to effectively learn complex, nonlinear mappings and enhance generalization across sharp orientation changes. To support untethered operation, we also introduce a compact, wireless electronic module for real-time EMF signal acquisition and transmission. Experimental results under dynamic conditions demonstrate a position error approximately 1.9 mm, an orientation error about 3.55°, and inference times under 1 ms. These findings validate the suitability of the system for real-time, closed-loop magnetic navigation in clinical environments without dependence on external computation.