Few-Shot Magnet Localization Using Sim-to-Real Transfer Learning

Magnet localization is widely used for real-time medical instrument tracking because it offers a radiation-free alternative to fluoroscopy. While physics-model-based tracking methods are standard, localization accuracy can suffer owing to imperfect physical assumptions and computational inefficiencies in iterative convergence. This paper proposes a learning-based method that uses a single, compact neural network to achieve efficient magnetic localization with low computation cost. The proposed method achieves few-shot generalization through transfer learning based on a simulation dataset. Mean magnet position and orientation errors are 1.66 mm and 2.15° in a 100×100×100mm workspace using only four 3-channel sensors, providing a 2× improvement compared to a magnetic model-based method. A real-time catheter tracking experiment demonstrates that the efficacy of the method for practical medical applications.