Page 219 - Artificial Intelligence for Computational Modeling of the Heart

192 Chapter 6 Additional clinical applications

a machine learning approach for investigating the relationship
between shape features and risk of rupture for aortic aneurysm.
This methodology has also been employed for accelerating CFD
computations: Hennigh [441] introduced a Deep Neural Network
architecture for reducing time and memory usage in Lattice Boltz-
mann flow computations, and Tompson et al. [442] used Convo-
lutional Neural Networks for accelerating fluid computations on
Eulerian grids.
In this section, we introduce a ML based pressure drop model
capable of accurately determining energy losses for a wide range
of flow conditions and anatomical CoA variations. Inspired by the
original work of Young and Tsai [438], where model parameters
were fitted to experimental data, the proposed ML based pressure
drop model was developed using a similar approach, but relying
on in silico data. A large number of 3D CFD computations were
performed on a set of synthetically generated aortic coarctation
anatomical models, and the resulting data was used to train a ML
based model for accurate CoA pressure drop prediction. A similar
approach was described in section 4.1 for estimating FFR in coro-
nary arteries. However in this case three-dimensional CFD simu-
lations were used instead of 1D, therefore fundamentally changing
the process of generating synthetic data and requiring a lot more
computationally intensive CFD simulations.

6.2.2 Methods
6.2.2.1 Generation of a synthetic training database
The process of generating synthetic CoA anatomical mod-
els consists in fitting a parameterized surface model to a set of
patient-specific models and using the fitted model to generate
synthetic CoA geometries.
In a first step, the given patient-specific anatomical models are
processed so that only the region of interest (descending aorta
containing the CoA segment) is retained and the resulting meshes
have the same orientation, scaling and position. The meshes are
then rotated such that the x axis becomes the direction of maxi-
mum extent and the y axis is the secondary direction of maximum
extent, as identified through principal component analysis on the
mesh points position. While mesh cutting is performed manually,
the rotations are performed automatically by using the computed
principal component vectors. Finally, the meshes are normalized
to ensure they are centered in the origin and cover a range be-
tween −1 and 1. The main purpose of rotation and normalization
is to maximize the overlap between different samples, since ro-
tation angles and absolute position in space are irrelevant to the