A Novel Approach to Quantitative Analysis of
Intravascular Optical Coherence Tomography Imaging

K Sihan1,2, C Botha2, F Post2, S de Winter1,
E Regar1, R Hamers1, N Bruining1

1Erasmus MC, Rotterdam, The Netherlands
2TU-Delft, Delft, The Netherlands
Abstract
Quantitative analysis on intracoronary optical
coherence tomography (OCT) image data (e.g. QOCT) is
currently performed by a time-consuming manual
contour tracing process in many individual OCT images
acquired during a pullback procedure (frame-based
method). In order to get a more efficient quantitative
analysis process and to investigate the possibilities to use
this contour information for development of an image-
based retrospective OCT-gating method, as a first step a
novel approach has been developed that exploits a full-
automated contour tracing method for coronary lumens
in OCT images without the need for intensive observer
related actions.
This study presents this new method, which was tested
on in-vivo acquired human coronary OCT image data of
10 randomly selected patients.
1. Introduction
The recent rapid developments in catheter based OCT
technology have led to great enthusiasm towards
applying this technique for the imaging of coronary artery
disease and applying it to the evaluation of new
therapeutic interventions. The major advantage of OCT
over the current reference method for intracoronary
imaging, intracoronary ultrasound (ICUS), is its
exceptionally high image resolution, close to that of the
golden standard of histopathology, i.e. lateral and axial
resolutions of 15 and 25 µm respectively. In contrast,
ICUS offers lateral and axial resolutions of 120 and 80
µm respectively [1]. To be able to select OCT for such
evaluation purposes, quantitative analysis tools are
imperative. The measurement accuracy of QOCT has
been validated and showed a good correlation both with
ex-vivo human coronary artery specimens as well as with
in-vivo acquired OCT data [2]. The latter is far more
tempting, since the imaging of the coronary vessel wall is
impossible with OCT due to the presence of blood. The
vessel must be occluded with a non-dilating balloon first
and than flushed with saline, before OCT imaging
becomes possible. Furthermore, induced motion artifacts
caused by the cardiac cycle do not only cause in-plane
artifacts (resulting in a sort of corkscrew appearance of
the lumen contour) but also cause the saw-tooth shaped
appearance of the coronary vessel wall. This limits the
possibility of using reconstructed longitudinal views of
the pullback OCT image dataset for contour detection,
which would be much faster than analysis of each
individual image [2]. Most acquired OCT image data sets
contain multiple hundreds of frames. To analyze all
frames is a time costly and tedious procedure currently
taking up to multiple hours. Therefore many observers
only analyze 1 frame per mm, randomly selected
throughout the cardiac phase. This could result in a
diminished accuracy for quantitative analysis.
To reduce the workload and to improve the accuracy
by eliminating possible observer-related induced
measurement inaccuracies, we investigated the possibility
and feasibility of developing a fully automatic lumen
contour detection approach. The newly developed
method has been tested on in-vivo data of 10 patients with
as reference method computer-assisted observer detected
contours.
2. Methods

2.1. OCT imaging

We used OCT images acquired with a commercially
available system (Lightlab Imaging, Westford, MA,
USA). The light source was a 1310-nm broadband super
luminescent diode with an output power in the range of
8.0 mW. The penetration depth in the tissue was
approximately 1.5 mm. The imaging catheter
(Imagewire™ Lightlab) had a maximum diameter of
0.019 inch and contained a single fiber optical core
within a transluminent sheath. The wire was pulled back
automatically at a speed of 1.0 mm/s while images were
acquired at a rate of 15 frames/s. The acquired studies
were saved in AVI format, and were then converted to
DICOM by in-house developed dedicated software.
ISSN 0276−6574 1089 Computers in Cardiology 2008;35:1089−1092.

Currently the OCT data still comes in different digital
formats.
The developed fully automated OCT contour
detection method contains the following steps:

2.2. Pre-processing

After importing the digitally stored OCT image data, a
pre-processing filtering step is applied to every individual
OCT image (Fig 1A). The primary goals of this step are
to reduce speckle noise inherent in OCT data and to take
care of gaps and shape irregularities in the blood-lumen
interface. The preprocessing step consists of a
combination of gaussian filtering and relative thresholds
to remove speckle noise, morphological closing to
remove gaps and contrast stretching to ensure that the
image is properly normalized.

2.3. Edge detection

The basic algorithm used to detect all coronary vessel
edges that can be found within a single OCT image,
including the lumen edge (or boundary), is the Canny
filter [3]. However, due to the rapid changes of the lumen
morphology between sequential OCT images, a
straightforward application of the Canny filter does not
directly lead to the desired result. Furthermore, there is a
large difference between the OCT datasets acquired from
different individuals. To tackle these problems, the
Canny filter is implemented iteratively using a binary
search, until the desired percentage of image pixels are
classified as edge pixels (Fig. 1). So the threshold of the
Canny filter is increased or decreased depending on the
amount of detected pixels.

2.3. Lumen edge selection

Not all edges detected by the Canny filter are on the
lumen contour. Some edges are from noise caused by the
catheter and some edges are from speckle noise. Even
after the applied pre-processing some speckle noise will
still be present in the images. Some parts of the contour
may also not be visible or are not pronounced enough to
produce an edge. In practice this is not such a problem
judging by the fact that most contours are closed and
clearly visible. The Canny filter also detects edges which
do not belong on the lumen contour. These edges are
removed using the dot product between the gradient
orientation and the catheter center. So every pixel where
this dot product is larger than a certain threshold is
removed (fig. 1D). However, some erroneous edges are
still present. The majority of these edges are short so a
threshold by length is used to remove these edges. The
subset of edges, which will be selected from the
remaining edges, is the one providing with the highest
quality score of all possible combinations. The quality
parameter is determined by the area-, length- (relative to
the area) and the gaps in the contour.

2.5. Post-processing

The pre-processing step sometimes pushes the edges
detected with the Canny filter away from the lumen
contour. In order to put the edges back on the lumen
contour a post-processing step is applied. This step
pushes contour points outwards based on the image
gradient magnitude and a local search of 5 pixels. The
contour is smoothed based on the radius length,
weighting neighbour coordinates applying the gradient
magnitude and the euclidean distance with a normal
distribution.



Figure 1. Panel A, shows an original OCT image. In
panel B, the result of the pre-processing step can be
appreciated and in panel C the edges diction results of the
Canny filter. Panel D, shows the lumen contour edge
selection and panel E the result after a threshold by edge
length. Finally the lumen contour is presented in panel F.
3. Results
The new method was evaluated on 10 in-vivo acquired
OCT pullback datasets of human coronary vessels. Every
individual image was analyzed by a human observer
using a computer-assisted software (Vessel analysis,
CURAD BV, Wijk bij Duurstede, The Netherlands) and
was used as reference method [2].
In these 10 datasets 1890 individual OCT images were
available. The manual analysis required several hours and
the automated method approximately 30-45 min (Fig. 2),
resulting in a processing time of between 2-5 sec per
1090

frame, depending on the image resolution. Of the 1890
automated detected lumen contours, 10 (0.5%) showed a
large deviation as compared to the manual contours. The
deviations were mostly caused by artifacts in the images,
such as reflections of the outer sheath of the catheter. For
possible necessary manual correction a graphical user
interface is available.
The mean lumen areas of the 10 cases as measured by
the human observer were 4.1±1.4 mm2 and with the
automated method 4.0±1.3 mm2; p=0.09 (calculated by a
student’s t-test), respectively.



Figure 2. Panel A, shows a longitudinal reconstruction
with the lumen contour presented on top of a total
pullback procedure. Panel B shows the lumen area results
of both methods.
Furthermore, linear regression analysis showed a good
correlation between the methods (Fig. 3).

Figure 3. Linear regression analysis of the 10 cases.
Finally, also a Bland-Altman analysis showed good
results without any outliers and with a mean relative
difference between the methods of –1.14±1.87% (Fig. 4).

4. Discussion and conclusions
This study shows that the fully automated lumen contour
detection of in-vivo human coronary OCT image data is
feasible by showing similar results as a human observer
without statistical significant differences.


Figure 4. Bland-Altman analysis of the 10 cases.
However, there is a small tendency, resulting in a
small systematic deviation, for the automated method to
position the contour more at the inside of the lumen than
on top of it, which seems to be the preference of the
human observers. Since this is a systematic deviation, it
plays no role in analysis of longitudinal studies and it was
also not statistically significant.
To our knowledge this is the first fully automated OCT
lumen contour detection method described. Although the
number of cases used for this feasibility study is a bit on
the low side, the results are promising. Although the fully
automated method needs a check for deviated frames, the
number of these false detected lumen contours is
relatively very low (e.g. 0.5%). Furthermore, the applied
quality score easily identifies those frames, so a complete
visual check of all analyzed frames is not necessary.
The advantages of a fully automated contour detection
method are the fast analysis time as compared to the
manual method, although computer-assisted, and of
course there are no inter- and intra-observer related
possible deviations. Furthermore, the results of a fully
automated lumen contour detection offer the possibility
to use this for retrospective image-based OCT-gating. As
can be appreciated in figures 2 and 5, the artifacts
induced by the cardiac cycle cause a saw-tooth shaped
appearance of the coronary vessel wall.
Since most cardiac imaging methods are gated it
would be desirable to have also the possibility to gate
OCT data since it improves not only the visual
appearance (e.g. a smooth representation of the vessel
wall providing a better matching with other imaging
modalities as ICUS or multi-slice computed tomography
(MSCT)) but also improves measurement accuracy [4].
Since OCT is becoming a popular choice for first-in-man
trials for the evaluation of new therapeutic interventions,
1091

accuracy is of utmost importance.



Figure 5. A 3D-plot of all full-automated detected lumen
contours of 1 of the analyzed cases.

Previous studies have shown that OCT and QOCT are
valuable additional imaging tools to the currently applied
reference method for intravascular coronary imaging,
ICUS. The advantages of OCT have been mentioned in
the introduction. However, the current disadvantages are
the necessary temporal closure of the vessel to be able to
flush it with saline before the vessel wall can be imaged,
and the low penetration depth limiting the visualization of
the outer vessel borders in areas with severe amounts of
plaque and finally cardiac motion induced artifacts.
Retrospective image-based gating of which the current
study was the first step could possibly solve the problem
of motion artifacts [5].
For the first two problems, new systems are currently
being developed and are in a research phase. They make
use of the optical frequency domain imaging (OFDI)
technique, allowing acquisition of images at the very high
frame rate of approximately 125 frames/s. By pulling the
catheter back at speeds of 15-25 mm/s at this high frame
rate, the vessel does not need to be occluded anymore.
However, these new advantages also bring new
disadvantages. One of them is that in this way acquisition
of images is performed during a few complete cardiac
cycles bringing together images acquired at diastole and
systole and everything in between. With the coronary
lumen area changing on average 10% during the cardiac
cycle this could lead to deviated results. This could limit
the application of the OFDI technique in longitudinal
atherosclerosis progression-regression trials. The current
OCT technique combined with the possible application of
gating could therefore provide a better platform when
performing multi-modality imaging studies comparing
the OCT image data against gated ICUS and gated MSCT
data.
The present study shows that fully automated lumen
contour detection in OCT images is feasible with a small
number of frames showing easily correctible artifacts.
The described method could be the first step towards a
retrospective image-based OCT-gating method.

References
[1] Regar E, Schaar JA, Mont E, Virmani R, Serruys PW.
Optical coherence tomography. Cardiovascular radiation
medicine. Oct-Dec 2003;4(4):198-204.
[2] Tanimoto S, Rodriguez-Granillo G, Barlis P, de Winter S,
Bruining N, Hamers R, Knappen M, Verheye S, Serruys
PW, Regar E. A novel approach for quantitative analysis of
intracoronary optical coherence tomography: high inter-
observer agreement with computer-assisted contour
detection. Catheter Cardiovasc Interv. Aug 1
2008;72(2):228-235.
[3] Canny J. A computational approach to edge detection.
IEEE transactions on pattern analysis and machine
intelligence. 1986;8:679-714.
[4] Bruining N, Tanimoto S, Otsuka M, Weustink A, Ligthart
J, de Winter S, Van Mieghem CA, Nieman K, de Feijter
PJ, van Domburg R, Serruys PW. Quantitative multi-
modality imaging analysis of a bioabsorbable poly-l-lactic
acid stent design in the acute phase: a comparison between
2- and 3D-QCA, QCU and QMSCT-CA. Eurointervention.
2008(August):285-291.
[5] De Winter SA, Hamers R, Degertekin M, Tanabe K,
Lemos PA, Serruys PW, Roelandt JR, Bruining N.
Retrospective image-based gating of intracoronary
ultrasound images for improved quantitative analysis: The
intelligate method. Catheter Cardiovasc Interv. Jan
2004;61(1):84-94.

Address for correspondence:

Nico Bruining
Erasmus MC/Room BA-571
Dr. Molewaterplein 40
3015 GD Rotterdam
The Netherlands
E-mail: n.bruining@erasmusmc.nl
1092