Physics-informed neural networks for patient-specific predictions in the cardiovascular system

Seminario del Dipartimento di Matematica
6 febbraio 2024
Orario di inizio 
PovoZero - Via Sommarive 14, Povo (Trento)
Aula Seminari (Povo 0)
Organizzato da: 
Dipartimento di Matematica
Comunità universitaria
Comunità studentesca UniTrento
Ingresso libero
Dott. Simone Pezzuto
Staff Dipartimento di Matematica
Francisco Sahli Costabal (Pontificia Universidad Católica de Chile- Santiago, CHILE)


In the cardiovascular system, we typically obtain indirect measurement of the quantities that are useful for diagnosis or understanding diseases. For instance, in cardiac electrophysiology, we can obtain electroanatomical maps, which show the arrival times of the electrical waves in the tissue. However, we are actually interested in the underlying structure of the tissue and observe if there any anomalies that may lead to malfunction. To solve this issue, we have pioneered the use of physics-informed neural network to infer the fiber orientation and conduction velocities in atria, which is something that cannot be imaged with any other method. This methodology not only relies on the available data, coming from electroanatomical maps, but also incorporates physical knowledge, which in this case corresponds to the dynamics of wave propagation. Combining physics and data through neural networks, we can infer hidden quantities in the atria, such as the fiber architecture. We also used this technique to infer the motion of the heart from cine MR images. Even though these images show how the heart is moving, actually extracting the motion to compute important clinical metrics such as strain is a challenging problem. We tackled this problem again with physics-informed neural networks, achieving state of the art performance in a public benchmark. Here, we incorporated knowledge of the mechanics of cardiac tissue to improve the predictions. Finally, we also present a methodology to infer patient-specific hemodynamic parameters from aortic flow measurements coming from phase contrast magnetic resonance images using physics informed neural networks.