Evaluation and Comparison of Textural Feature Representation for the

Detection of Early Stage Cancer in Endoscopy

Arnaud A. A. Setio

, Fons van der Sommen

, Svitlana Zinger

, Erik J. Schoon

and Peter H. N. de With

Dept. of Electrical Engineering, Eindhoven University of Technology,

Den Dolech 2, 5612 AZ, Eindhoven, The Netherlands

Dept. of Gastroenterology and Hepatology, Catharina Hospital,

Michelangelolaan 2, 5623 EJ, Eindhoven, The Netherlands

Keywords:

HD Endoscopy, Cancer Detection, Esophageal Irregularities, Co-occurance Matrix, Texture Spectrum

Histogram (TSH), Rotation Invariant Uniform Local Binary Patterns (RIULBP), Gabor Features.

Abstract:

Esophageal cancer is the fastest rising type of cancer in the Western world. The novel technology of High

Deﬁnition (HD) endoscopy enables physicians to ﬁnd texture patterns related to early cancer. It encourages

the development of a Computer-Aided Decision (CAD) system in order to help physicians with faster identi-

ﬁcation of early cancer and decrease the miss rate. However, an appropriate texture feature extraction, which

is needed for classiﬁcation, has not been studied yet. In this paper, we compare several techniques for tex-

ture feature extraction, including co-occurrence matrix features, LBP and Gabor features and evaluate their

performance in detecting early stage cancer in HD endoscopic images. In order to exploit more image char-

acteristics, we introduce an efﬁcient combination of the texture and color features. Furthermore, we add a

speciﬁc preprocessing step designed for endoscopy images, which improves the classiﬁcation accuracy. After

reducing the feature dimensionality using Principal Component Analysis (PCA), we classify selected features

with a Support Vector Machine (SVM). The experimental results validated by an expert gastroenterologist

show that the proposed feature extraction is promising and reaches a classiﬁcation accuracy up to 96.48%.

1 INTRODUCTION

Esophageal cancer is the fastest rising type of cancer

in the Western world. The majority of the patients are

diagnosed in a late stadium, in which the survival rate

is only about 2-20% (Howlader et al., 2012). With

the availability of HD endoscopes, specialist physi-

cians are able to identify early stages of esophageal

cancer, by looking for subtle changes in color and

texture (Kara et al., 2010). However, these subtle

indicators of early cancer are easily overlooked. As

a consequence, identiﬁcation of early cancer in the

esophageal tissue still requires signiﬁcant effort and

experience. A supporting system that selects the visu-

ally informative areas and forwards them to the physi-

cian for further analysis and judgment would be of

high relevance and help for the physician in identify-

ing early stage cancer in the esophagus. In this study,

we present the ﬁrst steps towards the development

of such a system. In comparative studies, various

methodologies have been applied for the detection

of anomalies in images obtained by a Wireless Cap-

sule Endoscope (WCE). They are classiﬁed into three

domains: spatial-domain features, frequency-domain

features and high-level features (Liedlgruber and Uhl,

2011). Spatial-domain features, for instance RIULBP

(Ojala et al., 2002), or co-occurrence texture features

(Haralick et al., 1973), are the most commonly used

feature types.

In studies that focus on frequency-domain fea-

tures, Gabor ﬁlters have shown to be successful

for the detection of intestinal juices (Vilarino et al.,

2006). Feature properties characterized at multiple

scales, using for instance a wavelet transform, also

show a better performance compared to most tradi-

tional spatial-domain features (Unser, 1995). High-

level features obtained using Canny edge detection,

are suitable for the real-time image processing sys-

tem due to the short processing time (Kang and Do-

raiswami, 2003).

However, there is no study yet on feature extrac-

tion using the novel technology of HD endoscopy. In

this paper, we compare several texture feature extrac-

tion techniques in order to ﬁnd the best technique

238

A. A. Setio A., van der Sommen F., Zinger S., Schoon E. and de With P..

Evaluation and Comparison of Textural Feature Representation for the Detection of Early Stage Cancer in Endoscopy.

DOI: 10.5220/0004204502380243

In Proceedings of the International Conference on Computer Vision Theory and Applications (VISAPP-2013), pages 238-243

ISBN: 978-989-8565-47-1

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

for early cancer detection. Our research also aims

at ﬁnding the best combination of features in order

to increase the classiﬁcation accuracy. Afterwards,

we perform clinical validation and analysis of the ob-

tained result. Moreover, we add a speciﬁc HD prepro-

cessing step to optimize classiﬁcation accuracy.

2 METHODS

In this study, we evaluate texture feature extraction

techniques for ﬁnding early cancer using HD endo-

scopic images. Since we want to be able to localize

the early cancers, we segment the image into small

square tiles prior to further analysis. Using images

annotated by an expert physician, we divide this set

of tiles into a training and a test set. Section 3 offers

more detailed discussion about this.

In order to evaluate the feature performance, we

use the algorithm depicted in Figure 1. During the

ﬁrst step, the input tiles are preprocessed to remove

irrelevant tiles or textures. In the second stage, tex-

ture features are extracted using the feature extractor

designed speciﬁcally for endoscopic images. In addi-

tion, we also propose the usage of combined feature

vector in order to improve the classiﬁcation accuracy.

After reducing the feature dimensions using PCA, the

obtained features are used for training or classiﬁcation

using SVM in the last stage.

In this section, we present features which have

been successfully used to describe texture and adjust

them for ﬁnding patterns associated with early cancer.

The section also explains how we reduce the dimen-

sions of the feature vectors, combine the feature vec-

tors into one vector, and then perform classiﬁcation.

Figure 1: Block Diagram of the Cancer Detection System.

2.1 Pre-processing of HD Images

Endoscopic images contain areas that may consider-

ably disturb texture feature analysis. These areas in-

clude specular reﬂection and irrelevant texture pat-

terns, which are not informative for cancer detection.

In general, the dichromatic plane is often used to

represent the specular reﬂection. However, the use

of the dichromatic plane is not efﬁcient in correcting

highly textured images, since it is designed for uni-

form color areas (Saint-Pierre et al., 2011). In other

studies, the specular reﬂections in endoscopic im-

ages are corrected using linear skew (Tchoulack et al.,

2008) or digital inpainting (Saint-Pierre et al., 2011).

These approaches estimate the values of the images

affected by specular reﬂection for visual processing

only. However, the corrected areas still contain irrele-

vant information because the textures are not real. For

this reason, we propose to detect and discard the tiles

containing specular reﬂection.

The parts of the image containing specular reﬂec-

tion of the endoscopic light source are usually much

brighter than the relevant parts of the image. To detect

these regions, we convert the tiles to grayscale and

apply an empirically determined upper threshold of

T = 220 to every pixel of the tile. If 0.1% of the pix-

els in a tile exceeds this threshold, the tile will be re-

jected for further processing. To reduce the inﬂuence

of texture that is not informative for cancer detection,

we propose the usage of a median ﬁlter. Compared to

other smoothing ﬁlters, it has a better performance in

removing noise, while preserving edges. Since edges

are of crucial importance for the texture of the image,

we have adopted such a 5 × 5 median ﬁlter. This win-

dow size is based on visual evaluation of experimental

results.

2.2 Feature Extraction

From the tiles obtained in the preprocessing step, rel-

evant features are extracted, using several techniques.

These features are then comparatively studied in or-

der to ﬁnd an optimal of their combination for cancer

detection. In spatial domain, we use seven features

based on the co-occurrence matrix, Texture Spectrum

Histogram (TSH), Local Binary Patterns (LBP), His-

togram of Oriented Gradients (HOG), and Dominant

Neighbors Structure (DNS). In the frequency domain,

we employ two Gabor-based features and Fourier fea-

tures. In addition, we also consider adding color in-

formation to the combined feature vector, since the

early cancer leads to both subtle color and texture dif-

ferences. Let us now provide more detail on each of

the applied features.

2.2.1 Co-occurrence based Texture Features

A gray-level co-occurrence matrix M

is deﬁned as

(i, j) =

∑

p=1

∑

q=1











1 I

(p,q) = i ∧

(p + d

,q + d

) = j

0 otherwise,

(1)

where I

is the m × n image in color plane c, p and

q are the pixel location, d = [d

] is the displace-

ment vector and i and j are the intensities of the corre-

EvaluationandComparisonofTexturalFeatureRepresentationfortheDetectionofEarlyStageCancerinEndoscopy

239

sponding pixel (Haralick et al., 1973). From these co-

occurrence matrices, statistical features are extracted.

Based on experimental results with the training set,

we select 7 features that perform best for endoscopic

images, namely: homogeneity, contrast, energy, en-

tropy, dissimilarity, correlation, and variance.

Applying Haralick’s algorithm to our case, 32

bins offer sufﬁcient information. We propose

exponential displacement vectors D

(k) = 2

for

k = 0,1,... ,

log

(min(m,n))

and angles θ =

,45

,90

,135

, to account for local as well as dis-

tant pixel relations.

2.2.2 Texture Spectrum Histogram (TSH)

As a textural counterpart of the well-known His-

togram of Color, the Texture Spectrum Histogram of-

fers information on the distribution of texture in an

image, based on small units of texture, called Texture

Units (TUs) (Wang and He, 1990). A TU is repre-

sented by 8 elements, which are neighbors of center

pixel g

in a 3 × 3 window. Each of the neighboring

elements g

is mapped to one of three possible values

(0, 1, 2) by











0 g

< g

1 g

= g

2 g

> g

p = 1, 2, ... 8. (2)

A texture unit TU is then computed by

i, j

∑

p=1

× 3

p−1

, (3)

where i, j deﬁne the pixel position of g

. Parameter

i, j

is calculated for all pixels of each tile. The oc-

currence histogram of TU is used as a feature vector.

2.2.3 Rotation Invariant Local Binary Patterns

Rotation Invariant Uniform Local Binary Patterns

(RIULBP) is an efﬁcient texture operator based on

a set of circular neighbors (Ojala et al., 2002). The

neighbors set consists of P elements lying on a circle

of radius R (Fig. 2). The pixel value of neighbors that

are not exactly in the center of pixels are estimated by

interpolation. Afterwards, RIULBP is computed by

LBP

riu2

P,R

(

∑

p=1

s(g

− g

) U (LBP

P,R

) ≤ 2,

P + 1 otherwise,

(4)

where U (LBP

P,R

) is the number of the spatial transi-

tion in the neighbors set and s(x) is the sign function.

To add more information on the textural features,

the contrast of local image texture is considered by

locally observing the values of individual pixels and

Figure 2: Circularly symmetric neighbor sets for different

values of P,R.

deriving mean and variance values from them. There-

fore, we calculate a rotation invariant variance mea-

sure (VAR) as

VAR

P,R

P−1

∑

p=0

− µ)

, (5)

where µ is the mean value of all neighbors lying on the

same circle. Since LBP

riu2

P,R

and VAR

P,R

complement

each other, the joint combination of LBP

riu2

P,R

/VAR

P,R

is more accurate in describing the image texture.

2.2.4 Histogram of Oriented Gradients (HOG)

HOGs are a modiﬁcation of Scale-Invariant Fea-

ture Transform (SIFT) descriptors (Dalal and Triggs,

2005). This method is based on evaluating a local his-

togram of image gradient orientations in a dense grid.

For better invariance to illumination and contrast, the

gradient strengths are locally normalized.

2.2.5 Dominant Neighbors Structure (DNS)

DNS is obtained by generating an estimated global

map representing the measured intensity similarity

between any given image pixel and its surrounding

neighbors within a certain window (Khellah, 2011).

In order to compute the intensity similarity, several

pixels i are used as the center of the search windows

. The distances between pixels i and several neigh-

boring pixels j on R concentric circles of various radii

are calculated using Euclidean distance d. Further-

more, the average value of the obtained neighborhood

structures is computed for all chosen pixels i.

2.2.6 Gabor Features

As a tool for local time-frequency analysis, the Gabor

wavelet is one of the important methods for texture

feature extraction (Zhang and Ma, 2007). It repre-

sents the response of cortical cells of human visual

system devoted to the processing of visual signals. In

our study we employ scale and rotation invariant Ga-

bor features.

VISAPP2013-InternationalConferenceonComputerVisionTheoryandApplications

240

The Gabor wavelet transform is described as the

convolutional operation of input image f (x,y) with

the complex conjugate of the Gabor function. The

transform can be written as

(x,y) =

∑

f (x − s, y −t)ψ

∗

(s,t), (6)

where s and t are the ﬁlter mask size, p = 0, 1,...,P−1

and q = 0,1,...,Q − 1 are the scale and direction val-

ues, respectively, and ψ

∗

is the complex conjugate of

the Gabor function. We use a dyadic function to de-

sign the Gabor ﬁlter. For each combination of p and q,

we calculate the mean µ

and standard deviation σ

of the ﬁltered image. These are used for the feature

vector.

In order to make the feature vectors invariant

against rotation, we employ two approaches: DFT-

based (Lahajnar and Kovacic, 2003) and the Circular

Shift (CS) (Ng et al., 2005). In DFT-based process-

ing, we compute the Discrete Fourier Transform of

the original feature vector for each scale. In the CS

approach, the feature vector is re-oriented based on

the dominant orientation d. The dominant orientation

d is shifted to the ﬁrst element of feature vector.

2.2.7 Fourier Features

Another method to capture texture, using the Fourier

feature vector, has been used to classify celiac disease

from duodenal images (Vecsei et al., 2008). In this ap-

proach, the Fourier domain image is calculated from

the input image using the Fourier transform. After-

wards, multiple ring-shaped ﬁlters are applied to the

center of the Fourier spectrum of each color channel

in order to differentiate the frequency characteristic in

several frequency regions.

For our purpose, this approach is adapted as fol-

lows. We use uniform width, W = R/Q, of rings

where R is the maximum width of the Fourier domain

image and Q denotes the number of rings. From each

ring, we calculate the mean and standard deviation,

which are used in the feature vectors.

2.3 Combined Feature Vector

In general, studies on irregularity detection in endo-

scopic images are typically based on single features

only. It results in non-optimal texture classiﬁcation,

since only one aspect of the irregularity is taken into

account, e.g. only color or only texture. To improve

the classiﬁcation accuracy, we propose to combine

our feature vectors for texture with simple color im-

age statistics, namely the sample mean and sample

variance per color plane. This should account for the

subtle difference in color that early cancerous tissue

shows.

Focusing on combining features in different com-

plementary domains, next to the combination of color

and texture features, we evaluate the combination of

spatial- and frequency-domain texture features. Each

feature extraction method produces a feature vector,

to F

, where M stands for the number of fea-

ture extraction methods used in the combined algo-

rithm. The individual feature vectors are concatenated

to generate the combined feature vector F

2.4 Dimensionality Reduction

In order to remove any redundancy and reduce the

dimensionality of the feature vector set, we employ

Principal Component Analysis (PCA). This technique

rotates the data such that the ﬁrst dimension has the

highest variance and the last dimension has the low-

est variance. Then only the ﬁrst D

∗

dimensions of

the rotated data are used for the reduced-dimensional

data. We select D

∗

by taking the ﬁrst N dimensions

for which the sum of the singular values is 0.95 of the

total sum of the singular values.

2.5 Classiﬁcation

SVM is a classiﬁcation method for two-group clas-

siﬁcation problems (Cortes and Vapnik, 1995). Re-

cently, it has gained considerable attention due to the

excellent performance in producing accurate and ro-

bust classiﬁcation results.

In this paper, SVM is implemented using the soft-

ware LIBSVM (Chang and Lin, 2011). An SVM clas-

siﬁer depends on cost parameter C, kernel function

K (x

) and training data D

. In order to ﬁnd the

optimal C and K (x

), we use a grid search and

10-fold cross-validation in order to evaluate the per-

formance parameters. We employ the Radial Basis

Function (RBF) kernel and determine the RBF width

γ in the grid search described above.

3 CLINICAL VALIDATION

For clinical evaluation of the proposed algorithm, we

have gathered 50 RGB endoscopic images with a res-

olution of 1600 × 1200 pixels for each class, being

‘tumorous’ and ‘normal tissue’. From each class, the

images are split into non-overlapping tiles with pixel

dimensions 25 × 25, 50 × 50, or 75 × 75. Afterwards,

600 tiles are selected for the dataset. We have used

40% of the total dataset for training and the other 60%

EvaluationandComparisonofTexturalFeatureRepresentationfortheDetectionofEarlyStageCancerinEndoscopy

241

Table 1: Classiﬁcation performance (%) of tumorous tissue in RGB color space. Best accuracy of each tile size is in bold.

Features

25 × 25 50 × 50 75 × 75

acc sen spe acc sen spe acc sen spe

Homogeneity 94.25 95.93 92.53 87.58 82.31 91.94 91.05 90.94 91.16

Contrast 94.25 93.90 94.62 90.20 84.48 94.93 92.77 90.59 94.90

Energy 94.60 97.29 91.84 87.58 86.28 88.66 90.71 90.94 90.48

Dissimilarity 94.17 95.59 92.71 89.22 85.20 92.54 92.43 91.29 93.54

Entropy 94.77 95.93 93.58 88.24 81.95 93.43 92.25 92.33 92.18

Variance 94.43 94.41 94.44 88.40 80.51 94.93 93.29 91.99 94.56

Correlation 92.62 93.39 91.84 89.22 87.36 90.75 91.57 91.99 91.16

TSH 86.36 86.95 85.76 80.39 73.29 86.27 83.65 77.70 89.46

HOG 90.65 91.36 89.93 88.24 80.51 94.63 88.12 89.90 86.39

LBP

riu2

P,R

84.13 84.24 84.03 85.46 80.14 89.85 84.51 83.62 85.37

VAR

P,R

93.57 95.59 91.49 88.89 85.20 91.94 90.88 92.68 89.12

LBP

riu2

P,R

/VAR

P,R

94.00 95.08 92.88 90.03 90.25 89.85 92.60 95.47 89.80

DNS 94.00 93.05 94.97 88.24 81.95 93.43 91.57 91.29 91.84

Gabor (DFT) 96.48 97.46 95.49 92.81 88.09 96.72 93.12 91.29 94.90

Gabor (CS) 96.40 97.63 95.14 92.16 87.73 95.82 94.15 91.99 96.26

Fourier 94.25 95.25 93.23 89.05 84.48 92.84 89.50 89.90 89.12

for testing. We tuned all classiﬁer and feature param-

eters using the training set. Table 1 shows the clas-

siﬁcation performance of our classiﬁer on the testing

sets. We have measured the performance using statis-

tical measures: accuracy (acc), sensitivity (sen) and

speciﬁcity (spe).

According to Table 1, the rotation and scale invari-

ant Gabor features perform better in terms of classiﬁ-

cation accuracy compared to the other methods. It is

due to the multi-resolution analysis at different scales

and directions, which is not available in other fea-

tures. The algorithm complexity of the Gabor feature

extraction is lower than the complexity of other fea-

tures. Therefore, we conclude that Gabor features are

suitable for real-time implementation in a further de-

velopment. Based on our experiment, the best classi-

ﬁcation accuracy for Gabor features is obtained using

P = 2 and Q = 8.

Since the Gabor features show the best perfor-

mance, we have tried to improve the classiﬁcation by

combining them with other features. After perform-

ing a number of experiments, we conclude that the

best results are obtained by combining Gabor features

with the proposed color information, i.e. the sample

mean and sample variance. Figure 3 shows a compar-

ison between using Gabor features alone and a Gabor

extension with color features, where the latter clearly

improves the classiﬁcation performance. This is in

line with our expectations, since early cancer is char-

acterized by differences in texture and color.

As an interesting visual result, Figure 4 illustrates

an original HD endoscopic image, the expert image

and an image where the tiles detected by the algo-

Figure 3: Accuracy measured of the test set: in blue

(darker color)-classiﬁcation result for Gabor texture alone;

in green-classiﬁcation for a combined feature vector (Gabor

and color information), for various tile sizes.

rithm are delineated. The expert image is created by

the expert gastroenterologist. Our detection result is

obtained by using the Gabor features with color infor-

mation on 50 × 50 tiles. From Figure 4, we can ob-

serve that our algorithm delivers a promising result,

which is very similar to the ground truth.

4 CONCLUSIONS

We have presented a novel benchmark for texture fea-

ture analysis applied to HD endoscopic images of the

esophagus for enhancing early-stage cancer detection.

This study is a step forward to a CAD system for the

real-time diagnosis based on HD endoscopic images.

We have studied several texture feature approaches

and concluded that Gabor multi-resolution analysis,

such as with Gabor features, provides better results

VISAPP2013-InternationalConferenceonComputerVisionTheoryandApplications

242

Figure 4: Original Image (left), Ground Truth Image (center), and Output Image (right).

than other features. It is suited to describe irregular

textures, associated with early cancer, containing var-

ious directions and scale. For further improving the

classiﬁcation accuracy, we propose a combined fea-

ture vector based on incorporating color information

and the insertion of a speciﬁc HD preprocessing step.

In the latter step we remove for example specular re-

ﬂections that normally confuse image analysis. For

efﬁcient classiﬁcation, PCA and SVM are employed

to reduce the dimensionality and to classify the fea-

ture vectors. Our proposed methodology achieves a

classiﬁcation accuracy up to 96.48%.

Future work should focus on detecting more sub-

tle abnormalities (earlier stages of cancer) and more

advanced pre-processing for specular reﬂection re-

moval. Novel post-processing can also be imple-

mented to provide a better image visualization.

REFERENCES

Chang, C.-C. and Lin, C.-J. (2011). LIBSVM: A library

for support vector machines. ACM Transactions on

Intelligent Systems and Technology, 2:27:1–27:27.

Cortes, C. and Vapnik, V. (1995). Support-vector networks.

Machine Learning, 20:273–297.

Dalal, N. and Triggs, B. (2005). Histograms of oriented gra-

dients for human detection. In Computer Vision and

Pattern Recognition, 2005. CVPR 2005. IEEE Com-

puter Society Conference on, pages 886 –893 vol. 1.

Haralick, R. M., Shanmugam, K., and Dinstein, I. (1973).

Textural Features for Image Classiﬁcation. Sys-

tems, Man and Cybernetics, IEEE Transactions on,

3(6):610 – 621.

Howlader, N., Noone, A., Krapcho, M., Neyman, N.,

Aminou, R., Altekruse, S., Kosary, C., Ruhl, J., Tat-

alovich, Z., Cho, H., Mariotto, A., Eisner, M., Lewis,

D., Chen, H., EJ, F., and Cronin, K. (2012). Seer can-

cer statistics review.

Kang, J. and Doraiswami, R. (2003). Real-time image

processing system for endoscopic applications. In

Electrical and Computer Engineering, 2003. IEEE

CCECE 2003. Canadian Conference on, volume 3.

Kara, M. A., Curvers, W. L., and Bergman, J. J. (2010). Ad-

vanced Endoscopic Imaging in Barrett’s Esophagus.

Techniques in Gastrointestinal Endoscopy, 12(2):82–

89.

Khellah, F. (2011). Texture classiﬁcation using domi-

nant neighborhood structure. Image Processing, IEEE

Transactions on, 20(11):3270 –3279.

Lahajnar, F. and Kovacic, S. (2003). Rotation-invariant tex-

ture classiﬁcation. Pattern Recognition Letters, 24.

Liedlgruber, M. and Uhl, A. (2011). Computer-aided deci-

sion support systems for endoscopy in the gastroin-

testinal tract: A review. Biomedical Engineering,

IEEE Reviews in, 4:73 –88.

Ng, C., Lu, G., and Zhang, D. (2005). Performance study

of gabor ﬁlters and rotation invariant gabor ﬁlters.

In Multimedia Modelling Conference, 2005. MMM

2005. Proceedings of the 11th International.

Ojala, T., Pietik

ainen, M., and Maenpaa, T. (2002). Mul-

tiresolution Gray-Scale and Rotation Invariant Tex-

ture Classiﬁcation with Local Binary Patterns. IEEE

Transactions on Pattern Analysis and Machine Intel-

ligence, 24(7):971–987.

Saint-Pierre, C.-A., Boisvert, J., Grimard, G., and Cheriet,

F. (2011). Detection and correction of specular re-

ﬂections for automatic surgical tool segmentation in

thoracoscopic images. Mach. Vision Appl., 22(1).

Tchoulack, S., Pierre Langlois, J., and Cheriet, F. (2008).

A video stream processor for real-time detection and

correction of specular reﬂections in endoscopic im-

ages. In Circuits and Systems and TAISA Conference,

2008. NEWCAS-TAISA 2008. 2008 Joint 6th Interna-

tional IEEE Northeast Workshop on, pages 49 –52.

Unser, M. (1995). Texture classiﬁcation and segmentation

using wavelet frames. Image Processing, IEEE Trans-

actions on, 4(11):1549 –1560.

Vecsei, A., Fuhrmann, T., and Uhl, A. (2008). Towards

automated diagnosis of celiac disease by computer-

assisted classiﬁcation of duodenal imagery. In Ad-

vances in Medical, Signal and Information Process-

ing, 2008. 4th IET International Conference on.

Vilarino, F., Spyridonos, P., Pujol, O., Vitria, J., and Radeva,

P. (2006). Automatic Detection of Intestinal Juices in

Wireless Capsule Video Endoscopy. In Pattern Recog-

nition (ICPR’06), 18th International Conference on,

pages 719–722, Hong Kong. IEEE.

Wang, L. and He, D.-C. (1990). Texture Classiﬁcation using

Texture Spectrum. Pattern Recognition, 23(8).

Zhang, G. and Ma, Z.-M. (2007). Texture feature extraction

and description using gabor wavelet in content-based

medical image retrieval. In Wavelet Analysis and Pat-

tern Recognition, 2007. International Conference on,

volume 1, pages 169 –173.

EvaluationandComparisonofTexturalFeatureRepresentationfortheDetectionofEarlyStageCancerinEndoscopy

243