ECGI without Geometry: A Deep Learning Based Estimation of Heart Surface Potentials

We developed a deep learning framework to estimate heart surface potentials solely from body surface potentials and a torso geometry, hence omitting information about the heart geometry. The framework is based on image-to-image translation and introduces three components: transformation of 3D torso and heart geometries into corresponding standard 2D representations, and development of a customized deep learning model based on Pix2Pix network. Using 11 healthy subjects and 29 idiopathic ventricular fibrillation (IVF) patients, the framework achieved an average mean absolute error (MAE) for the heart surface potential maps (HSPMs) of 0.012 ± 0.011, and an average similarity index measure (SSIM) of 0.984 ± 0.026. For the concatenated electrograms (EGMs), the average MAE was 0.004 ± 0.004, and the average Pearson correlation coefficient (PCC) 0.643 ± 0.352. The absolute mean of the time differences between estimated and reference activation and recovery times was 6.048 ± 5.188 ms, and 18.768 ± 17.299 ms, respectively. These results demonstrate performance comparable to standard ECGI without the need for CT/MRI, indicating potential clinical applications for this framework.