Page 148 - Artificial Intelligence for Computational Modeling of the Heart

120 Chapter 4 Data-driven reduction of cardiac models

vides results in real-time. Such supervised machine learning (ML)
algorithms are routinely employed in medical imaging applica-
tions, e.g. organ segmentation [332]. Moreover, machine learning
models can also be employed to reproduce the behavior of non-
linear computational models [37,333].
Herein, a machine learning model for near real-time FFR pre-
diction is presented [334]. The main goal was to obtain a method
providing results which are statistically indiscernible from those
obtained with the CFD based approach. To build a large enough
training database synthetically generated pathological coronary
anatomical models are employed. Cycle-averaged pressures, and
derived FFR, are computed at each centerline location of each
synthetic tree using a reduced-order CFD model [315]. Anatomi-
cal features are then determined at each location of each coronary
tree, and paired with the corresponding CFD-based FFR values to
build the training database. Finally, a deep neural network (DNN)
is trained to predict FFR. The resulting model computed FFR in
2.4 ± 0.44 seconds, whereas the CFD model required 196.3 ± 78.5
seconds, both on a standard desktop computer with a 3.4 GHz In-
tel i7 8-core processor. Ideally, the machine learning model should
be trained on CFD results obtained for patient-specific anatom-
ical models. Since the costs and the effort for setting up such a
large training database are prohibitive, purely synthetically gen-
erated anatomical models of the coronary tree are used during
the training phase, and generated the corresponding ground truth
FFR values using the CFD model. The generation of the synthetic
database follows a parameterization approach relying on anatom-
ical features of the arterial coronary tree, which enables the gen-
eration of various configurations like serial stenoses, three-vessel
disease, bifurcation stenoses, diffuse stenoses or rare pathologies.
In the following, cFFR CFD refers to the FFR values computed
with the CFD based model, while cFFR ML refers to the FFR val-
ues computed with the ML based model. Verification and vali-
dation are performed in three steps: (i) cFFR ML vs. cFFR CFD
on synthetic anatomical models, (ii) cFFR ML vs. cFFR CFD on
patient-specific anatomical models, (iii) cFFR ML vs. invasive FFR
on patient-specific anatomical models.

4.1.2 Methods
In the following a framework employed for machine learning
based FFR computation in coronary arteries is described:
• Generation of synthetic coronary arterial trees
• CFD based approach for computing FFR in coronary arteries
• Feature definition and extraction for generating the mapping
between the coronary anatomy and FFR