Learning-based Magnetic Localization Model Enhanced by Regularization Techniques

In recent years, deep learning-based magnetic localization algorithms have been extensively studied in the medical field. They can be used to localize medical devices that penetrate the human body during minimally invasive surgeries, thereby improving surgical safety. However, noise from magnetic sensors and the inherent nonlinearity of magnetic fields can easily affect the positioning accuracy of the algorithms. Therefore, this study focuses on investigating the impact of different regularization techniques on the training performance of multilayer perceptrons (MLPs) in magnetic localization tasks, where high-quality training data are generated based on the magnetic dipole model. It systematically analyzes the effects of Dropout and Batch Normalization on MLP stability, generalization ability, and prediction accuracy, aiming to enhance the model's robustness in noisy environment. Our comparative analysis reveals a trade-off between the two methods: Dropout provides lightweight regularization effects but with limited scope; while Batch Normalization delivers more comprehensive improvements, it relies on batch statistics and may affect stability in small-batch scenarios. Experimental results show that Dropout reduces angle error by 9.3% while preserving positioning accuracy, and Batch Normalization achieves a 20.6% improvement in directional prediction accuracy alongside a 7.6% reduction in positioning error (RMSE). This study emphasizes the importance of regularization techniques in MLP training under noisy environments and suggests that future research could explore adaptive hybrid regularization strategies as a direction.