Page 145 - Artificial Intelligence for Computational Modeling of the Heart

P. 145

Data-driven reduction of cardiac

models

a
a
a,b
Lucian Mihai Itu , Felix Meister , Puneet Sharma ,
Tiziano Passerini a
a Siemens Healthineers, Princeton, NJ, United States. Friedrich-Alexander
b
University Erlangen-Nuremberg, Erlangen, Germany
Physiological models of cardiac biology and biomechanics allow
high ﬁdelity representation and simulation of the mechanisms
governing the functions of the heart at the cellular, tissue, or-
gan and system level. The underlying mathematical formulation
of the models typically requires complex numerical approxima-
tion strategies and high computational costs, as discussed in the
ﬁrst part of this book. This often induces a trade-off between the
descriptive power of the model and its utility. Simplifying assump-
tions can reduce the complexity of the numerical approximation,
and the time required to compute a solution; at the expense of the
capability of the model to describe the complex, multi-scale dy-
namics typical of biological systems, and to capture inter-patient
variability. Different approaches are described in this chapter to
reduce the computational complexity of physiological models,
while still providing accurate and detailed representation of the
physical phenomena of interest. These approaches are described
through three use cases, covering cardiac hemodynamics, electro-
physiology and biomechanics.
Coronary artery disease (CAD) is the most prevalent cardiovas-
cular disease (CVD). Plaque builds up in the coronary arteries and
limits the ﬂow to the myocardium, especially when the demand
is increased (exercise, stress), potentially leading to myocardial
infarction, or even death. For the non-invasive evaluation of the
functional signiﬁcance of coronary artery disease, presented in
section 4.1, the explicit computation of patient-speciﬁc hemody-
namics in the coronary tree is replaced altogether with a model
trained using deep learning, providing statistically indistinguish-
able results, and enabling drastic reduction in the compute time.
The deep learning model expresses the relation between anatom-
ical features of the coronary tree that can be automatically ex-
tracted from medical images, and the patient-speciﬁc blood pres-

140 141 142 143 144 145 146 147 148 149 150