J. Biomedical Science and Engineering, 2010, 3, 1143-1145
doi:10.4236/jbise.2010.312148 Published Online December 2010 (http://www.SciRP.org/journal/jbise/ JBiSE ).
Published Online December 2010 in SciRes. http://www.scirp.org/journal/JBiSE
Detection of bleeding patterns in WCE video using TV-Retinex
Ming Li
Department of Computer Science, Huazhong University of Science and Technology, Wuhan, China
Received 12 Octorber 2010; revised 18 Octorber 2010; accepted 20 Octorber 2010.
ABSTRACT
The Retinex theory is used to deal with the removal
of unfavorable illumination effects from images. In
this paper, we present the Retinex theory for bleeding
detection in wireless capsule endoscopy (WCE). This
processing is quite appropriate to refresh old bleed-
ing region and bleeding region in shadow. A novel
total variation model (TV-Retinex) is proposed to
solve the Retinex problem quickly; also a support
vector machine is employed for classification. Ex-
perimental results demonstrate the efficacy of the
proposed method.
Keywords: Wireless Capsule Endoscopy; Bleeding
Detection; Retinex Theory; Total Variation
1. INTRODUCTION
Wireless Capsule Endoscopy (WCE) is a novel tool for
visualizing abnormalities in the gastrointestinal tract,
widely used to replace traditional endoscopy diagnosis,
with advantages of the capability for reaching even the
duodenum and small intestine. About 50,000 images
are obtained during an exam. Usually, these images are
reviewed in a form of a video at sp eeds between 5 to 40
frames per second, and the time spent reading the video
varies between 45 to 180 minutes. This long review
time was usually reported as the major weakness of
capsule video endoscopy.
To reduce the assessment time, many methods have
been developed for automatic or semi-automatic detec-
tion of certain type of abnormal images in WCE video.
The bleeding region is one of the most widely used
features for its simple and association with many dis-
eases. The first software tool to detect bleeding images
is provided by the manufacturer; however, sensitivity
and specificity of this system are reported to be only
21.5% and 41.8%, respectively [1]. Recently, Li and
Meng [2] proposed a method using Tchebichef poly-
nomials and hue/saturation/intensity (HSI) color space,
combined with uniform local binary pattern (LBP) to
classify a dataset contain 3600 bleeding patches and
3600 non-bleeding patches. The reported specificity
and sensitivity are 93.2% and 91.6%. However, some
shadow patches are removed in their experiments.
2. TV-RETINEX
In recent years, with the development of color con-
stancy theory, enhancement algorithms based on the
Retinex theory have become a hot spot of research. The
most applied algorithms of them are single-scale Reti-
nex algorithm (SSR) [3] and multi-scale Retinex algo-
rithm (MSR) [4]. Recently, Kimmel [5] introduces a
varia t ional mo d e l fo r th e R eti nex problem.
 


2
22
Minimize:
Subject to: , and ,0 on
Elllsl sdxdy
ls ln





(1)
where

is the support of the image, its bound-
ary, and 
n

is the normal to the boundary.

and

are free non-negative real parameters. A Projected
Normalized Steepest Descent (PNSD) algorithm is de-
veloped to solve this problem. For image of size
250*250 pixels, it is usually needs 3-5 seconds to
process. Although the above model produces good re-
sults, the slow computation makes it unsuitable to deal
with large number of images. To speed up computation,
we modify the model and adopt a more efficient algo-
rithm for computing. The proposed model is defined as
follow:
 
2
22
2
22
s.t. , ,0 on
ElTVllsl s
ls ln


 
 (2)
We use instead of ()TV l2
2
ldxdy
.This change
allows us to use a more efficient split Bregman algo-
rithm [6] for calculation. To apply Bregman splitting,
we first replace l

by an auxiliary variable dl


to decoupling the L1 and L2 terms in (2).This yields
the constrained problem

M. Li et al. / J. Biomedical Science and Engineering 3 (2010) 1143-1145 1144

2
2
*12 2
argmin 22
l
ld lsls


l
(3)
where .To solve this constrained problem, we
convert it to an unconstrained problem by introducing a
quadratic penalty function:
d


2
** 12
(, )
22
2
2
,argmin 2
22
ld
ldd ls
lsd l



 

(4)
Although this problem can be solved using an alter-
nating minimization process, it suffers from slowly
convergence rate for large

, which is needed to en-
force the constraint exactly. To avoid this
difficulty, the Split Bregman approach enforces the
constraint using a Bregman iteration technique [7].
dr
Hence, in this work we added a Bregman vector
inside of the quadratic penalty function. Then it be-
come a sequence of unconstrained problems defined by,
k
b


2
11 12
(, )
2 2
2
2
,argmin
2
22
kk
ld
k
ldd ls
lsd lb




 
(5)
11kkkk
bbld

 1
kk
(6)
Now, the unconstrained problem (5) can be solved
using a simple alternating minimization scheme as fol-
low:
The first step is to minimize with respect to . This
is a differentiable optimization problem and the solu-
tion is obtained by solving
l
  

1k
I
lIsbd  

 . (7)
The system is strictly diagonally dominant, so we
can use a fast iterative algorithm to get approximate
solutions to it, such as Gauss-Seidel method. We next
minimize (5) with respect to d. This optimization
problem is element-wise decoupled. We can explicitly
compute it using shrinkage operators.

11
,1
kk
dshrinklb

k


 (8)
where
  ,max
x
shrink xx
x
 ,0

 . (9)
When we put all of these pieces together, and the
constraint , we get the following very simple, yet
efficient algorithm:
ls



1
1Retinex
11
111
while do
max,,,
,1
kk
kk
kkk
kkkk
ll
lGSsdb
dshrinklb
bbld









 
k
s

Here, we use to denote one
sweep of the Gauss-Seidel formula. For a 250*250 size
image, our method needs at most 0.15 seconds on a
personal computer, and the enhancement results are
nearly same as shown in Figure 1.

Retinex ,,
kk
GSs db
3. EXPERIMENTAL RESULTS
In order to compare the classification performance on
original images and images after TV-Retinex, we built
two datasets. First we select bleeding patches and
non-bleeding patches on original images by experts to
construct an original dataset, and then choose the cor-
responding patches from the enhanced images to form
another dataset. The processing of selecting bleeding
patch es on or iginal imag e s is as fo llo w :
We select 150 images containing bleeding regions of
various kinds from a WCE image database. Another
1000 images containing no bleeding regions are col-
lected from 10 patients’ video segmen ts, each for about
100 images. The size of each image is 280×280. For
every bleed ing image, the bleed ing region is mar ked by
a medical expert with a year of experiences in WCE
images analysis. We then divide every image into small
patches with 20*20 pixels, as illustrated in Figure 2,
since larger siz e may make s ome very sma ll symptoms
undetected; every patch seems homogeneous for color
feature, and finally it is much more robust than a single
pixel.
The patch in the bleeding region is marked as ab-
normal, the patch outside the bleeding region as normal,
and the patches come across the region edge and out-
side the region of circle are discarded as shown in Fig-
ure 2.
The patches in the 1000 non-bleeding images are all
marked as normal. Finally, we get 4,000 bleeding
patches and 50,000 non-bleeding patches for the ex-
Figure 1. The results of TV-Retinex in comparison with Ron’s
method. The first row shows the original images, the second
row shows the results of TV-Retinex, and the third row shows
the results of Ron’s method.

M. Li et al. / J. Biomedical Science and Engineering 3 (2010) 1143-1145 1145
4. CONCLUSION
We propose a new fast calculated color image enhance-
ment method called TV-Retinex for the detection of
bleeding in WCE video. The experimental results show
that the image enhancement processing is helpful fo r the
detection.
REFERENCES
[1] Ftancis, R.D. (2004) Sensitivity and specificity of the red
blood identification (RBIS) in video capsule endoscopy.
The 3rd International Conference on Capsule Endoscopy,
Miami, FL.
Figure 2. Patches chosen from one WCE image.
Table 1. Classification results with and withour TV-Retinex. [2] Baopu, L. (2009) Computer-aided detection of bleeding
regions for capsule endoscopy images, IEEE Trans. on
Biomedical Engineering, 56, 1032-1039.
Before TV-Retinex After TV-Retinex
Color
Space sensitivity specificity sensitivity specificity
RGB 90.2% 98.5% 96.3% 99.3%
HSI 88.3% 98.5% 95.3% 99.1%
RGB+HSI 89.3% 98.6% 96.6% 99.5%
[3] Jobson, D.J., Rahman, Z., and Woodell, G. A. (1997) Prop-
erties and performance of a center/surround retinex.
IEEE Transactions on Image Processing, 6, 451-462.
[4] Jobson, D.J., Rahman, Z., and Woodell, G.A. (1997) A
multiscale retinex for bridging the gap between color
images and the human observation of scenes. IEEE
Transactions on Image Processing, 6, 965-976.
[5] Kimmel, R., Elad, M., Shaked, D., Keshet, R. and Sobel, I.
(2003) Variational framework for retinex. International
Journal of Computer Vision, 52, 7-23.
periments. The unbalance proportion of normal and
abnormal samples is in line with the actual situation,
although it will reduce the accuracy of the system
compared with balanced proportion of samples [8]. [6] Goldstein, T., and Osher S. (2008) The split bregman method for L1 regularized problems. UCLA CAM Report,
08-29.
We use th e average color of each patch as its f eature
vector. (R,G,B),(H,S,I) and (R,G,B,H,S,I) feature vec-
tors are tested on a radial basis kernel SVM using 10
fold cross-validation. We tabulated the best classifica-
tion results of each feature vector on the two databases
in Table 1. As can be seen in Tabl es 1, the TV-Retinex
processing significantly improves the classification in
all the three different feature sets, and the RGB is better
than HSI color space.
[7] Bregman, L. (1967) The relaxation method of finding the
common points of convex sets and its application to the
solution of problems in convex optimization. USSR
Computational Mathematics and Mathematical Physics,
7, 200-217.
[8] Weiss, G.M. (2001) The effect of class distribution on
classifier learning: An empirical study. Technical Report
ML-TR-44, Rutgers University, New Jersey.
JBiSE