Page 229 - Artificial Intelligence for Computational Modeling of the Heart

202 Chapter 6 Additional clinical applications

database is an important prerequisite for the successful develop-
ment of such a model. In an ideal scenario, the training database
would consist of thousands of anatomical models extracted from
medical images, accounting for the variability of aortic coarcta-
tions across different patient populations, and the corresponding
invasive pressure measurements of each narrowing. From a prac-
tical point of view, establishing such a large database would be
prohibitively expensive and time-consuming.
To address this issue, we reused the concepts introduced in
[334]: a training database was generated consisting of synthetic
aortic geometries, and corresponding pressure drop values com-
puted from a CFD algorithm. The synthetic database is parameter-
ized by the morphological features of the aortic configuration, al-
lowing for a proper sampling of a relatively large range of anatomi-
cal configurations, addressing eccentricity, bulging, curvature, etc.
After training, the machine learning algorithm encodes the corre-
lation between the set of chosen geometric features and the quan-
tity of interest, herein pressure drop, predicted by the validated
3D CFD model. The considered geometric features are highly spe-
cialized for pressure loss prediction, being related to viscosity and
turbulence (as present in the original Young-Tsai model), convec-
tion, eccentricity and bulging. As can be observed in Table 6.3,
after model optimization all seven pressure loss coefficients are
non-zero, indicating that the CoA pressure drop is significantly in-
fluenced by each of the considered features.
The results here presented show the superiority of the pro-
posed deep-learning based pressure drop model compared with a
traditional one, in approximating pressure drop values produced
by a 3D CFD model in a database of synthetically generated vascu-
lar geometries representative of CoA anatomies. The performance
of the learning based pressure drop model could be further im-
proved by increasing the number of synthetic anatomical mod-
els and the corresponding number of computations performed
with the 3D CFD model. The set of features could also be fur-
ther expanded to include other characteristics of the coarctation,
e.g. aorta curvature in the CoA region, non-circular vessel cross-
section, etc.
It should be noted that the presented approach is generic with
respect to the CFD model used for generating the ground truth in-
formation. For this work, we used a 3D CFD model based on the
Lattice Boltzmann Method. On the other hand, the accuracy of the
pressure drop model compared to invasively measured pressures
will depend on the accuracy of the CFD model used in the train-
ing phase, which was not addressed in this discussion. Since the
proposed machine learning based pressure drop model learns the