Page 156 - Artificial Intelligence for Computational Modeling of the Heart

128 Chapter 4 Data-driven reduction of cardiac models

4.1.3 Results
The proposed methodology was evaluated in three steps, de-
scribed in the following.

4.1.3.1 Validation on synthetic anatomical models
The synthetic datasets were split randomly using a ratio of 5:1
into training and validation datasets. Hence, 2000 out of the 12 000
synthetically generated anatomical models were used for an initial
validation step. CFD based FFR was compared with ML based FFR
at all centerline locations and obtained an almost perfect correla-
tion results (ρ = 0.9998, p <0.001). The mean difference between
cFFR ML and cFFR CFD was 0.0008 (no systematic bias). When
considering CFD based FFR as ground truth, and employing 0.8
as binary cut-off value, ML based FFR has an accuracy of 99.7%.

4.1.3.2 Validation on patient specific anatomical models
A dataset based on 87 patients was further employed for vali-
dation. An anatomical model was reconstructed for each patient
from the CT data, using centerline detection and lumen border
segmentation, as described in [322,328]. As for the synthetic eval-
uation dataset, the correlation between cFFR ML and cFFR CFD ,
at all coronary centerline locations, was excellent (ρ = 0.9994,
p <0.001) and the Bland–Altman analysis revealed no systematic
bias: mean difference −0.00081 ± 0.0039 (Fig. 4.7,Fig. 4.8). The
mean computation time was 2.4 ± 0.44 seconds for ML based
FFR, and 196.3 ± 78.5 seconds for CFD based FFR, both on a
standard desktop computer (3.4 GHz Intel i7 8-core CPU). For
a more detailed analysis the individual centerline locations have
been grouped into 5 bins, based on the cFFR CFD values: 0.0–0.6;
0.6–0.7; 0.7–0.8; 0.8–0.9; 0.9–1.0. The bins were defined either
globally or separately for the three main coronary branches: LAD
(Left Anterior Descending), LCx (Left Circumflex) and RCA (Right
Coronary Artery). The results of this detailed comparison between
cFFR ML and cFFR CFD analysis are displayed in Table 4.2: overall
the agreement between cFFR CFD and cFFR ML is very high for all
locations and all bins.

4.1.3.3 Validation against invasively measured FFR
FFR was measured invasively for 125 stenoses in the above
mentioned dataset containing 87 patients. The exact measure-
ment location was set either around 2 cm distal from the steno-
sis [317] or, if available, extracted from the coronary angiogra-
phies displaying the pressure wire. All lesions with FFR ≤ 0.80
were considered hemodynamically significant [327] (38 out of 125