Page 14 - Artificial Intelligence for Computational Modeling of the Heart

List of figures xiii

displayed in green (mid gray in print version) and the ground-truth
in yellow (light gray in print version); and finally, a 3D rendering of
the corresponding triangulated surface mesh for the aortic root
(both computed using MSDL). 104
Fig. 3.5 Schematic overview of the multi-scale image navigation paradigm
based on multi-scale deep reinforcement learning. 107
Fig. 3.6 Left: The LV-center (1), the anterior/posterior RV-insertion points
(2)/(3) and the RV-extreme point (4) in a short-axis cardiac MR
image. Middle: The mitral septal annulus (1) and the mitral lateral
annulus points (2) in a cardiac ultrasound image. Right: The center
of the aortic root in a frontal slice of a 3D-CT scan. 109
Fig. 3.7 Visualization of the considered cardiac and vascular landmarks. 110
Fig. 3.8 Segmentation masks for heart isolation computed with a deep
neural network. 112
Fig. 4.1 Overall workflow of the proposed method. 121
Fig. 4.2 Three stage approach for generating synthetic coronary
geometries: (A) Define coronary tree skeleton, (B) Define healthy
coronary anatomy, (C) Define stenoses. 123
Fig. 4.3 Multiscale model of the systemic and coronary arterial circulation. 124
Fig. 4.4 Deep neural network model employed for computing cFFR ML :
fully connected architecture with four hidden layers. 125
Fig. 4.5 Features describing a stenosis. 126
Fig. 4.6 Examples of hemodynamic interdependence between coronary
branches. 127
Fig. 4.7 Scatterplot of cFFR ML and cFFR CFD (correlation = 0.9994). 129
Fig. 4.8 Bland–Altman analysis plot comparing cFFR ML and
cFFR CFD : no systematic bias was found (95% limits of
agreement, −0.0085 to 0.0067). 129
Fig. 4.9 (A) Scatterplot of cFFR CFD and invasive FFR (correlation =
0.725); (B) Scatterplot of cFFR ML and invasive FFR (correlation
= 0.729). 131
Fig. 4.10 Bland–Altman analysis plot comparing cFFR CFD and cFFR ML
vs. invasive FFR (cFFR CFD 95% limits of agreement, −0.159 to
0.207; cFFR ML 95% limits of agreement, −0.159 to 0.206). 131
Fig. 4.11 Receiver operating characteristic (ROC) curves of cFFR CFD
and cFFR ML for 125 lesions. 132
Fig. 4.12 Case example of a patient-specific coronary tree with cFFR ML
color coded coronary tree: invasive FFR = 0.71 and cFFR ML =
0.68. 132
Fig. 4.13 Diagram of the ionic channels in the CRN atrial cell model
(source: CellML, https:/ /
models.cellml.org/ exposure/ 0e03bbe01606be5811691f9d5de10b65). 138
Fig. 4.14 Left: Action potential of the CRN model with parameters as in
[359]; Middle: Current I ion –I Na ; Right: Current I Na . 140
Fig. 4.15 Samples with SD=0.3 by different number of parameters. 145
Fig. 4.16 Goodness of reconstruction on testing data using PCA and LLE. 146
Fig. 4.17 Modes of variation of the action potential profile produced by the
CRN model, estimated by PCA components. 147
Fig. 4.18 Regression on V peak and V rest : PLSR (1st column), MARS