An Automated System for 3D Segmentation of CT Angiograms - Visual Computing: Scientific Visualization and Imaging Systems - page 107

Fig. 5.9 Construction of the ground truth surface from manually delineated vessel boundaries.
a The boundary points of the left coronary artery constructed using the centreline and the
corresponding radius information obtained via manual annotation, b The outer surface of the
artery reconstructed based on the boundary points shown in (a)

d H ð X ; Y Þ ¼max f sup
x 2 X

y 2 Y d ð x ; y Þ ; sup
y 2 Y

inf

x 2 X d ð x ; y Þg

inf ð 5 : 11 Þ

where X, Y are the vertices of the mesh surfaces of the arteries corresponding to
the segmentation results and the ground truth, respectively, and d(x, y) measures
the Euclidean distance between points x and y belonging to vertices X and Y. The
mesh surface of the arteries was obtained by extracting the isosurface of the binary
volume obtained from the segmentation/manual delineation, using the marching
cube algorithm.
To demonstrate the efficiency of incorporating local intensity features into
active contour functional, we compare the performance of the proposed method
with the work proposed by [ 16 ], which utilises global intensity information alone,
in extraction of the arteries in clinical images. In Figs. 5.10 and 5.11 , Tables 5.1 ,
5.2 and 5.3 , we present the comparison of the resulting segmentation obtained
using the proposed technique and Yang et al.'s method [ 16 ], with respect to the
ground truth data. The initial surface for the active contour models was obtained
through the application of a Hessian-based vessel enhancement filter [ 27 ]. The
tuning parameters of both of the two techniques were empirically determined from
a training set, which consisted of three CT studies randomly selected from the
available datasets. Specifically, for the proposed approach, we set u = 0.2, v = 0.1
and the radius of localised kernel was set to 7 voxels. The proposed approach was
implemented in MATLAB (R2010b) on a standard specification PC (Dell Preci-
sion T3500, Inter(R) Xeon(R) CPU at 2.67 GHz), and the average execution time
was found to be 80 s for extraction of the entire coronary trees. Yang et al.'s
algorithm, on the other hand, requires roughly 45 s to carry out the same process.
As shown in Table 5.1 , the mean TP rate and OM metric for the proposed
method were found to be 91.1 % and 0.776, respectively, which indicate that the
proposed method is able to correctly extract the major branches of the coronary
arteries (see Fig. 3.10 a-c). Meanwhile, the high values of the FP rate (39.2 % on
average) mean that the proposed method over-segments the arteries, as illustrated in
Fig. 5.11 , where the segmentation results were shown on the 2D axial image as
contours. In these axial images, the red contours represent the ground truth