Region-Selective Synthetic Data Injection for Data-Driven Magnetic Capsule Pose Estimation
Stevanus Darwin, Ayoung Hong
AI summary
Problem
Accurate wireless magnetic capsule pose estimation requires large-scale training data, but physical data acquisition is slow and limited in volume. This bottleneck restricts the effective workspace and hinders the development of robust data-driven estimators.
Approach
The method replaces real sensor measurements with outputs from a calibrated physics-based model in regions where the model is highly accurate, while retaining actual sensor data in lower-fidelity regions to train a neural network.
Key results
- X-axis position RMSE of 0.28 mm, matching purely data-driven baselines
- Y and Z-axis RMSE increases to 0.53 mm and 0.63 mm in physics-injected regions
- Establishes a 70 µT RMSE calibration threshold to reliably separate high- and low-fidelity regions
- Significantly reduces physical data acquisition time while expanding training workspace coverage
Why it matters
Provides a scalable, efficient training framework for magnetic capsule robotics that balances computational efficiency with safety-critical accuracy for minimally invasive procedures.
Abstract
Accurate Wireless Magnetic Capsule (WCE) pose estimation remains a challenge for advancing minimally inva- sive medical procedures, because the relationship between mag- netic sensor measurements and capsule pose is highly nonlinear and sensitive to noise and modeling errors, making large-scale training data essential for data-driven estimation. However, data acquisition itself remains a limiting factor, restricting both the volume of training data and the effective workspace of the system. To address this limitation, we propose a region-selective synthetic data injection strategy that generates additional data points using a calibrated physics-based model. In this strategy, regions with high model fidelity are replaced with physics- based data at arbitrary points, while regions with lower fidelity rely on sensor data, which provides a more accurate representation of the real system. Experimental results show that the proposed strategy achieves performance comparable to that of a purely data-driven model while significantly reducing the data acquisition burden.