Page 147 - Artificial Intelligence for Computational Modeling of the Heart

Chapter 4 Data-driven reduction of cardiac models 119

percentage diameter stenosis, determined during invasive coro-
nary angiography (ICA). The degree of reduction in lumen di-
ameter may be determined either visually or through computer-
assisted quantitative coronary angiography [314].
Non-invasive imaging techniques, like Coronary Computed
Tomography Angiography (CCTA), play nowadays an increasingly
important role for CAD diagnosis, prior to ICA. CCTA has been
shown to provide a high negative predictive value, but, by overesti-
mating the lesion severity [315,316], it also leads to a large number
false positives.
The purely anatomical CAD assessment, irrespective of the
medical imaging modality, does not correlate closely with the
functional significance of the coronary lesions. Hence, the diag-
nostic measure of Fractional Flow Reserve (FFR) has been pro-
posed as an alternative [317]. FFR is defined as the ratio of cycle-
averaged pressure distal to the stenosis and cycle-averaged aortic
pressure, both measured during hyperemia. It represents a sur-
rogate measure for the reduction of hyperemic flow caused by
the stenosis, and has been shown to lead to superior long-term
patient outcomes, when compared to the purely anatomical as-
sessment [318]. Although FFR is now the gold standard in inter-
national guidelines [319,320], it is still not used at a large scale,
mainly due to the higher costs and risks, and the longer procedure
times [321].
In view of these limitations, approaches based on compu-
tational fluid dynamics (CFD) have been introduced for deter-
mining FFR non-invasively from medical images acquired at rest
[322–329]. These models are run under patient-specific condi-
tions: the coronary anatomical model is reconstructed from med-
ical images, and physiological information is derived and em-
ployed for model personalization. The CFD models consist of
partial differential equations, which can be solved only numeri-
cally, leading to a very large number of algebraic equations. Thus,
the solution requires several hours in case of high-fidelity three-
dimensional models, and several minutes in case of reduced-
order models [315,330]. Due to the computationally intensive as-
pect of CFD models, and the time-consuming process required
for generating the anatomical model, they are not used for intra-
operative assessment and planning, where real-time performance
is required. Alternatively, artificial intelligence based solutions
may be employed, capable of providing results instantaneously. To
develop such solutions, a large training database is required, con-
taining input–output data pairs, where input data is represented
by the vascular geometry, and the output is FFR [331]. Once the
training phase has been completed, the online application pro-