Information Technology Reference
In-Depth Information
5.4 Experiments and Results
The proposed method was tested on real clinical datasets. Twelve coronary CT
volumes were acquired from St Thomas' and Guys' Hospitals, London, UK. Two
of them were obtained with a 16-slice CT scanner (Brilliance, Philips), and the
remaining volumes were acquired with a Philips ICT-256 workstation. The mean
size of the images is 512 9 512 9 286 with an average in-plane resolution of
0.40 mm 9 0.40 mm, and the mean voxel size in the z-axis is 0.42 mm. For each
CTA image, four major coronary arterial branches, namely, Right Coronary Artery
(RCA), Left Anterior Descending Artery (LAD), Left Circumflex artery (LCX)
and one large side branch of the coronaries were evaluated. To quantify the per-
formance of the resulting segmentation, four metrics were used to validate the
results, specifically:
TP ¼
N
B
\
N
R
N
R
FN ¼
N
R
N
B
\
N
R
N
R
;
ð
5
:
10
Þ
FP ¼
N
B
N
B
\
N
R
N
R
OM ¼ 2
N
B
\
N
R
N
B
þ
N
R
;
where the ground truth N
R
is a binary image with voxels labelled to one for the
object and zero for others, N
B
indicates the voxels, which are segmented as the
object by the aforementioned algorithms. TP, FN and FP denote the true positive,
false negative and false positive metrics, respectively. OM represents the over-
lapping metric defined in [
26
], which is close to 1, when the segmentation is well
matched to the reference ground truth and approaches zero when the results have
no similarity to the reference.
The ground truth data were obtained through manual delineation with the
assistance of interactive software, developed in our centre. To construct the ground
truth data, the CT volume was firstly resampled as a successive cross-sections,
which is perpendicular to the course of the arteries, as shown in Fig.
5.7
. Then, the
luminal area was manually annotated by trained biomedical engineering research
students (see Fig.
5.8
a). The software will fit the manually delineated curve to a
circle by solving the associated least squares problem (see Fig.
5.8
b). The software
then records the coordinates of the centre and the associated radius. In order to
ensure the correct generation of the ground truth data, the luminal area of the
artery is required to be specified by the observer at least every 3 mm. By doing so,
the ground truth data for each major branch of the arterial tree contains on average
48 central axis points and the associated radii, which takes approximately half an
hour to complete. Next, the centreline points were uniformly resampled with a
distance at 0.05 mm (roughly 0.1 voxel), and the associated radii were determined
via linear interpolation. To construct a closed surface of the ground truth data, we
firstly generate the boundary points of the artery based on the centreline and radius
information, which is depicted in Fig.
5.9
a. Then, the outer surface of the artery
can be reconstructed using the ball pivoting algorithm, as illustrated in Fig.
5.9
b.
Search WWH ::
Custom Search