Page 163 - Artificial Intelligence for Computational Modeling of the Heart

Chapter 4 Data-driven reduction of cardiac models 135

4.1.4.2 Limitations
Several limitations need to be mentioned in the context of the
study. As shown in the previous section, the deep learning model
was able to almost perfectly reproduce the output of the hemody-
namic model, and, thus the limitations are mainly defined based
on the latter. First, the patient-specific test dataset was not large
enough to allow for a generalization of the diagnostic statistics,
and subsequent clinical studies are required for further validation.
For example, herein, almost all FFR measurements had a value
between 0.7 and 1.0 (FFR < 0.7wasfoundinonly11lesions). Nev-
ertheless, the most challenging lesions in terms of diagnosis are
those with an FFR value close to the cut-off point of 0.8. A recent
study [347] indicated that for the threshold of 0.87 and 0.74 the
measures of NPV and PPV respectively were both larger than 90%.
Thus a hybrid revascularization strategy may be defined, whereas
a cFFR value greater than 0.87 may be used for PCI deferral, a cFFR
value smaller than 0.74 may be employed for treatment indica-
tion, whilst the remaining lesions may be assessed invasively. The
overall agreement with invasive FFR would in this case be higher
than 95% in terms of accuracy, and could potentially eliminate
the requirement for invasive assessment in around half of the pa-
tients. Secondly, the parameters in the CFD model are based on
physiological observations, which should be validated extensively.
For example, allometric scaling laws are employed to determine
the flow rate in each branch, the hyperemic response (in case of
microvascular dysfunction, pharmacological vasodilators gave a
limited effect, which in turn may lead to an overestimation of the
lesion severity), population average blood density and viscosity,
etc. Since the latter were set to constant values in the synthetic and
patient-specific hemodynamic computations, they have not influ-
enced the correlation between the blood flow and the machine
learning model. Nevertheless, these properties can be introduced
as additional input variables/features and accounted for using a
proper sampling in the training database. Furthermore, collateral
vessels were not taken into account. These can alter the blood
flow characteristics, especially if severe lesions are present. The
synthetic anatomical models represent many common patholog-
ical coronary configurations. Nevertheless, the 12 000 synthetic
geometries do not address all types of rare pathologies, like coro-
nary aneurysms or anomalous coronary artery origin. Moreover,
smooth stenotic shapes were employed which do not reflect the
noisy shape variations in stenoses reconstructed from medical
images. A future activity will focus on the generation of more re-
alistic stenotic shapes. While by employing the herein proposed
approach the cFFR values can be computed in near real-time