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

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Artificial Intelligence

Background:

• Electrocardiographic imaging (ECGI) is vital for non-invasive cardiac mapping.
• Current ECGI methods face challenges in accuracy and robustness.
• Accurate pacing site localization is critical for treating cardiac arrhythmias like premature ventricular contractions (PVCs).

Purpose of the Study:

• To develop an innovative approach combining deep learning with classical ECGI for enhanced reconstruction accuracy and robustness.
• To improve the precision of pacing site localization using ECGI.
• To validate the proposed method against existing ECGI techniques.

Main Methods:

• Proposed a novel method (PULSE) sequentially applying analytical solutions and a denoiser neural network.
• The denoiser learns cardiac potential priors directly from data, avoiding manual assumptions.
• Compared PULSE against Tikhonov regularization, Bayesian MAP estimation, and end-to-end learning.

Main Results:

• Achieved over 10% improvement in all metrics compared to Bayesian-MAP, end-to-end learning, and Tikhonov solutions.
• Demonstrated consistent performance across cardiac beats, reducing metric interquartile ranges by 60%.
• Maintained accuracy despite geometric variations, with median localization error consistently below 1cm.

Conclusions:

• The PULSE framework significantly enhances ECGI reconstruction accuracy and robustness.
• The method offers a promising advancement for pacing site localization in PVC research.
• PULSE is adaptable to classical ECGI methods, potentially augmenting clinical pipelines.