![](bg2.png)
structed pathway, since such percentage is a relative
score that requires a reference complete ring for an ac-
curate estimation. On the other hand, semi-automatic
techniques are infeasible in the operating room.
Developing accurate computer procedures for ex-
tracting the ring area from videobronchoscopic im-
ages would constitute a significant breakthrough in
the field. This is challenging task due to the large
variety of acquisition conditions and devices, which
include, among others, flexible and rigid optics, dif-
ferent video camera resolutions and digital compres-
sions. Besides, processing videos acquired at the op-
erating room adds the unpredicted presence of surgi-
cal devices (such as probe ends), as well as, illumina-
tion and camera position artifacts. These multiple im-
age artifacts together with a large diversity of anatom-
ical structures not belonging to the set of tracheal
rings are prone to drop the performance of methods
exclusively relying on image intensity (S´anchez et al.,
2011).
This paper introduces a model of tracheal rings
that combines their appearance and geometric fea-
tures in videobronchoscopic images in order to mini-
mize the impact of non-tracheal structures. In video-
bronchoscopic sequences, the trachea is described as
a tube in conical projection. The appearance of rings
in the image follows a ridge-valleypattern that is geo-
metrically characterized by its concentric disposition
and an increasing radial thickness. The ridge-valley
profile is modelled by a bank of Normalized Steer-
able Gaussian Filters (NSGF) in order to minimize
the impact of illumination and camera position vari-
ations. In order to account for the concentric dispo-
sition, images are transformed to the polar domain.
Finally, the increasing thickness is taken into account
by analysing along each radius the scale of the maxi-
mum response to NSGF.
We present experiments on videos acquired at
the operating room with different devices, both, with
rigid and flexible cases. Final results show that the
error rate is very close to inter-observer variability.
This validates our methodology as a further tool for
helping pulmonologist in assessing the percentage of
stenosed trachea.
Paper contents are: Section 2 explains our mod-
elling of tracheal rings in terms of their appearance in
images (Section 2.1) and geometric structure (Section
2.2). Section 3 describes the video data set and vali-
dation protocol used (3.1), as well as, numeric results
(3.2). Finally, Section 4 concludes the paper.
2 TRACHEA
GEOMETRIC-APPEARANCE
MODEL
The trachea is a tubular structure located in front
of the esophagus that connects the pharynx to the
beginning of the bronchial tree (known as the ca-
rina). There are about fifteen to twenty incomplete
C-shaped cartilaginous rings that reinforce the ante-
rior and lateral sides of the trachea. In videobroncho-
scopic images, the trachea appears as a tube in coni-
cal projection. If the camera is centered at the carina,
the conical projection of the trachea is given by a set
of concentric circles corresponding to tracheal rings.
Considerable deviation from the center causes rings
to collapse at a certain pathway point. In this case,
measurements are not reliable and, thus, these images
are usually discarded by experts. For this reason, we
will only consider images having the camera centered
at the carina for our further analysis on ring detection.
The left image in figure 1 shows the main anatomi-
cal structures that can be identified in a videobroncho-
scopic frame acquired in central projection. The dark
central spot shows the two main bronchi separated by
the carina where the tracheal rings are the concentric
bright-dark structures around it. The brighter parts
show to the C-shaped cartilage and the dark ones to
separating soft tissue. The illumination intensity is
not uniform due to variations in the incidence angle
between camera light and ring surface. We also ob-
serve an increasing radial thickness of cartilage and
separating tissue due to the perspective projection of
the video camera. The plots in figure 1 show the im-
age profile for the radial lines with origin at the carina
labelled L1 and L2 and shown in solid white in the
right image of figure 1. The main anatomical struc-
tures observed across the radial profile are indicated
in both images. It is worth noticing that the profile
along L1 illustrates the fact that bronchial rings at the
carina have a radial local ridge-valley profile similar
as the tracheal ring ones.
2.1 Ring Appearance Modelling
As already pointed out in section 2, tracheal rings
have a ridge-valley intensity profile in videobroncho-
scopic images transformed into gray level. Among
current ridge-valley detectors, we have chosen convo-
lution with a second derivative of an anisotropic ori-
ented gaussian kernel (Freeman and Adelson, 1991).
Oriented anisotropic gaussian kernels are given by:
G
Σ,θ
= G
(σ
x
,σ
y
),θ
=
1
(2π)σ
x
σ
y
e
−
˜x
2
2σ
2
x
+
˜y
2
2σ
2
y
(1)
VISAPP2013-InternationalConferenceonComputerVisionTheoryandApplications
154