Speaker
Description
Computational cardiology is based on the numerical solution of complex partial differential equations to model cardiac electrophysiology from non-invasive measurements. However, high-resolution simulations on anatomically realistic geometries are computationally expensive, whereas clinical practice demands rapid, interpretable, and application-oriented predictions. In this talk, we highlight recent developments in scientific machine learning for both forward and inverse cardiac problems, with a focus on operator learning and neural surrogate models.
We first describe operator learning strategies, in particular Fourier Neural Operators (FNOs) and Kernel Operator Learning (KOL), designed to learn the mapping from activation patterns in the physical domain to cardiac activation and repolarization times. These neural operators are trained on synthetic 2D and 3D domains as well as on a realistic left ventricle geometry. The learned operator for activation times aligns with the Eikonal formulation, while the operator predicting repolarization times has no explicit PDE analogue, showcasing the adaptability and expressiveness of data-driven operator learning.
We then focus on inverse cardiac problems, aiming at reconstructing ischemic areas and pacing sites from pseudo-ECG simulations. To this end, we employ an architecture inspired by Latent Dynamics Networks (LDNets), which serves as a fast neural surrogate of pseudo-ECG signals generated by the monodomain model. This approach enables efficient forward evaluations within an inverse learning framework on both 2D and 3D ventricular geometries.
These results illustrate how machine-learning-based surrogate and operator models can dramatically reduce computational costs for cardiac simulations and inverse reconstructions, paving the way toward clinically impactful applications.