Page 162 - Artificial Intelligence for Computational Modeling of the Heart

134 Chapter 4 Data-driven reduction of cardiac models

for each stenosis. However, both the costs and the effort to set up
such a large database are prohibitive. To mitigate this aspect, only
synthetically generated anatomical models of the coronary tree
were employed during the training phase, and generated the cor-
responding ground truth FFR values using a previously validated
blood flow model. The generation of the synthetic database fol-
lows a parameterization approach relying on anatomical features
of the arterial coronary tree, which enables the generation of vari-
ous configurations like serial stenoses, three-vessel disease, bifur-
cation stenoses, diffuse stenoses or rare pathologies. As a result of
the training process, the deep learning model encodes the map-
ping between the geometric characteristics chosen as features and
the measure of interest, here FFR, computed by the blood flow
model. The anatomical characteristics of the patient-specific data
used to define the test set were well within the ranges of values ob-
served in the synthetic training dataset. Given the generic nature
of the approach, other coronary hemodynamic measures of inter-
est, like rest PdPa [353], CFR (coronary flow reserve), IFR (instant
wave-free ratio) [354], BSR / HSR (basal / hyperemic stenosis resis-
tance) [355,356], wall shear stress [357], etc, may be employed as
ground truth during the training process. Moreover, the purely lu-
men based features defined herein may be augmented with other
features related to coronary hemodynamics, e.g. plaque charac-
teristics and patient clinical history, which are important not only
for the functional assessment of a lesion but also for its vulnerabil-
ity in time [358]. Depending on the quantity of interest, a different
blood flow model may be employed for defining the ground truth
values. Herein, a reduced-order blood model was used, but full-
order models may be employed since the time required to build
the training database does in general not represent an issue. We
note also that the diagnostic performance of cFFR ML is closely
linked to the diagnostic performance of the blood flow model
employed during the training process. If features are properly se-
lected and the datasets are large enough, the diagnostic statistics
of the machine learning model will be indiscernible from those
of the blood flow model. Since cFFR is computed at all center-
line locations in the anatomical model, virtual pullback curves
can be easily extracted, displaying the cFFR variation from a distal
location up to the coronary ostium. Pullback curves are specifi-
cally useful to distinguish between focal and diffuse lesions and to
evaluate the hemodynamic significance of serial lesions indepen-
dently.