Page 139 - Artificial Intelligence for Computational Modeling of the Heart

Chapter 3 Learning cardiac anatomy 111

Table 3.3 Accuracy at detecting different anatomical landmarks in incomplete 3D-CT scans. The
accuracy is measured in mm.

Landmark FPR FNR Mean STD Med.
Carotid Artery Merge 0% 0% 2.68 2.96 1.65
Basilar Artery Branch 0% 0% 1.75 2.17 1.16
Right Intracranial 0% 0% 1.88 1.65 1.54
Left Intracranial 0% 0% 2.52 1.90 2.31
Right Carotid Frontal 0% 0% 1.60 1.63 1.11
Left Carotid Frontal 0% 0% 1.07 0.99 0.69
Right Carotid Skull 0% 0% 1.44 1.51 1.07
Left Carotid Skull 0% 0% 1.03 1.01 0.74
Right Vertebral Artery 0% 0% 0.87 1.32 0.60
Left Vertebral Artery 0% 0% 1.24 0.95 1.10
Right Brachiocephalic Tr. 0% 0% 1.89 2.67 0.66
Left Brachiocephalic Tr. 0% 0% 1.92 2.77 0.67
Right Vertebral Artery C3 0% 0% 1.00 1.59 0.75
Left Vertebral Artery C3 0% 0% 1.00 1.26 0.67
Right Vertebral Artery C5 0% 0% 3.30 4.90 1.42
Left Vertebral Artery C5 0% 0% 1.98 3.90 0.77
Vertebral Art.–Basilar Art. 0% 0% 2.24 2.89 1.06
Right Subclavian–Vertebralis 0% 0% 4.68 4.39 2.93
Left Subclavian–Vertebralis 0% 0% 3.99 3.85 2.68
Left Common Carotid Art. Bif. 0% 0% 2.99 2.51 2.20
Brachiocephalic Art. Bif. 0% 0% 2.25 1.99 1.57
Left Subclavian Art. Bif. 0% 0% 2.55 2.76 1.50
Aortic Arch Center 0% 0% 1.96 1.22 1.72
Aortic Root 0% 0% 3.34 2.21 2.97
Celiac Trunk 0% 0% 2.36 1.88 1.97
Aortic Bifurcation 0% 0% 1.33 1.27 1.07
R. Proximal Common Iliac 0% 0% 2.25 2.54 1.27
L. Proximal Common Iliac 0% 0% 2.40 2.70 1.51
Renal Bifurcation 0% 0% 1.81 2.31 1.29

In terms of runtime, the reformulation of the detection as a
multi-scale navigation process enables an average detection time
per landmark of 52 milliseconds on CPU (Intel 8-core) and 28 mil-
liseconds on GPU (Nvidia Pascal).