Symmetry-based Independent Fully Convolutional Neural Networks for Magnetic Localization Using Sensor Array

Magnetic tracking offers a promising alternative to fluoroscopic imaging for localizing intra-body medical devices. Conventional approaches rely on mathematical magnetic field models combined with iterative algorithms. However, these methods are computationally intensive, sensitive to initial solutions, and prone to divergence. Although artificial neural networks have addressed these limitations, they typically require extensive training data, which is impractical in clinical settings. In this article, a convolutional neural network-based method is developed to address these challenges. To reduce network complexity, orientation and position estimations are decoupled into independent networks. Additionally, three transfer functions are derived to estimate the mirrored states of the magnet across the sensor array’s plane axes and origin. A transfer learning strategy is also implemented, using synthetic data for pretraining and experimental data for fine-tuning. The proposed approach is experimentally evaluated on a 4×4 magnetometer array and benchmarked againts conventional magnetic localization methods. The results demonstrate superiority of the proposed method in accuracy, and efficiency, achieving a localization accuracy of 0.7 ± 0.4 mm in position and 3.7 ± 2.1° in orientation, with an average execution time of 1.0 ± 0.07 ms.