PULSE: A DL-Assisted Physics-Based Approach to the Inverse Problem of Electrocardiography

Abstract

This study introduces an innovative approach combining deep-learning techniques with classical physics-based electrocardiographic imaging (ECGI) methods. Our objective is to enhance the accuracy and robustness of ECGI reconstructions. We reshape the optimization expression by splitting variables and formulating building blocks based on update expressions. Specifically, we propose a sequential application of analytical solutions and denoiser neural network blocks (PULSE). The denoiser learns the proximal operator associated with the prior distribution of cardiac potentials directly from data, avoiding hand-crafted assumptions about the distribution. The proposed method is compared with zero-order Tikhonov regularization, Bayesian MAP estimation, and an end-to-end learning technique. We achieved more than 10% improvement in all metrics over Bayesian-MAP, end-to-end learning, and Tikhonov solutions. The performance remained consistent throughout cardiac beats, resulting in a 60% reduction in the interquartile ranges of the reconstruction metrics. Geometric variations did not compromise accuracy, with a median localization error consistently below 1cm. Our framework, adaptable to classical methods, augments the clinical pipeline. Improving the accuracy and robustness of pacing site localization holds significant promise for premature ventricular contraction (PVC) research.