Projekt

Computational models of the heart are poised to assist clinical decision making, from triage (e.g. non-invasive fractional flow reserve) to individualized therapy selection (e.g. cardiac resynchronization therapy) to intervention guidance (e.g. ablation therapy of cardiac arrhythmias). These models simulate the complex, multi- scale biophysics involved in cardiac function.
Siemens Healthineers has developed a unique, end-to-end technology stack to go from medical data to virtual hearts that includes anatomical, electrophysiological, biomechanical and fluid-dynamics modeling. Although already fast, a speed up of one or more order of magnitude is still needed to bring this technology to the bedside: first, to be able to estimate the virtual heart model from medical data in minutes and not hours, as it currently does due to the large number of simulations over several heart beats needed for parameter estimation; and second to allow real- time, model-based guidance during interventions.
The objective of this work is to explore AI-based approaches to enable extremely-fast but accurate biophysical simulations. Building upon the universal approximation theorem, we will research the capability of neural networks to learn the underlying physical laws driving computational models and consequently whether they can circumvent the need for accurate numerical solvers like the finite element method, which is known to be computationally complex. In addition, we will explore other applications such as the discovery of governing laws that cannot be captured in a closed form and the full differentiability of neural-network-based physics solvers. Finally, the results of these investigations will be integrated into an existing prototype and validated clinically.