Layer-Conditioned Magnetic Localization With Residual-Weighted Fusion for Wireless Capsule Endoscopy Under Asymmetric Sensor Visibility and GI Motion

Magnetic-based localization is a promising non-invasive approach for tracking wireless capsule endoscopes (WCE), but its performance degrades sharply under involuntary gastrointestinal (GI) motion, asymmetric visibility, and magnetic saturation. Conventional single-layer systems, as well as multi-layer methods that treat all sensors as a single undifferentiated layer, become unstable once several sensors drift outside their effective operating range. Prior methods do not explicitly model asymmetric layer visibility as a failure mode, leaving a gap in robustness under dynamic GI displacement. This paper presents a dual-layer magnetic localization framework designed to address these limitations through layer-specific modeling and fusion. The proposed framework integrates layer-wise Levenberg–Marquardt estimation, grid-based initialization to mitigate local minima, and layer-specific calibration to compensate for systematic magnetic distortions. Each layer functions as an independent dipole-based estimator, and their outputs are combined using a residual-weighted fusion mechanism that suppresses unreliable layer contributions during abrupt GI displacement or partial saturation. Robotic-arm experiments using an N52 ( $10\times 12$ mm) magnet and an eight-sensor two-layer array demonstrate mean positional and angular errors of approximately 2.5 mm and 2°, even when one layer temporarily loses visibility — representing a 20–45% reduction in positional error compared to prior single-layer and layer-agnostic methods reporting 3–5 mm accuracy, using fewer sensors than comparable systems. To evaluate hardware generalizability, the experiments were repeated using a smaller N52 ( $5\times 5$ mm) magnet and high-resolution MMC5983MA sensors. The optimal array geometry remained unchanged, and the proposed layer-specific approach consistently outperformed both single-layer and layer-agnostic baselines. The proposed system-level integration of layer-conditioned estimation, residual-weighted fusion, and layer-specific calibration demonstrates that robust, motion-resilient WCE localization is achievable with only eight sensors, offering a practical and generalizable framework for clinical translation.