A Comparative Study of Texture Features for the

A Comparative Study of Texture Features
for the Discrimination of Gastric Polyps
in Endoscopic Video
D. Iakovidis1, D. Maroulis1, S.A. Karkanis2, A. Brokos1
1
University of Athens
Department of Informatics & Telecommunications
Realtime Systems & Image Processing Laboratory
2
Technological Educational of Lamia
Department of Informatics & Computer Technology
Gastric Cancer & Polyps
• Gastric Ca is the 2nd Ca-related cause of death
• Rarely alarming symptoms
• >40% appear as polyps
• Gastric polyps are visible tissue masses
protruding from the gastric mucosa
• Adenomatous polyps are usually precancerous
• Gastroscopy is a screening procedure with
which polyp growth can be prevented
Aim
Computer Science
Computer-Based Medical System (CBMS)
to support the detection of gastric polyps
• Increase endoscopists ability for polyp localization
• Reduction of the duration of the endoscopic
procedure
• Minimization of experts’ subjectivity
Previous Works
• Detection of gastric ulser using edge detection
(Kodama et al. 1988)
• Diagnosis of gastric carcinoma using epidemiological
data analysis
(Guvenir et al. 2004)
Previous Works
• Detection of colon polyps using texture analysis
1. Texture Spectrum Histogram (TS)
(Karkanis et al, 1999) (Kodogiannis et al, 2004)
2. Texture Spectrum & Color Histogram Statistics (TSCHS)
(Tjoa & Krishnan, 2003)
3. Color Wavelet Covariance (CWC)
(Karkanis et al, 2003)
4. Local Binary Patterns (LBP)
(Zheng et al, 2004)
Texture Spectrum Histogram
(Wang & He, 1990)
• Greylevel images
• 33 neighborhood thresholded in 3 levels
• V0 central pixel, Vi neighboring pixels, i =1, 2, …8
• Texture Unit TU = {E1, E2,…, E8}
• Totally 38 = 6561 possible TUs
• Feature vectors formed by the NTU distribution
0 if

Ei  1 if
2 if

Vi  V0
Vi  V0
Vi  V0
8
NTU   Ei  3
i 1
i 1
Local Binary Pattern Histogram
(Ojala, 1998)
• Greylevel images
• Inspired by the Texture Spectrum method
• 33 neighborhood thresholded in 2 levels
• Totally 28 = 256 possible TUs
• Feature vectors formed by the NTU distribution
0 if
'
Ei  
1 if
Vi  V0
Vi  V0
8
LBP   Ei'  2i 1
i 1
Texture Spectrum and Color
Histogram Statistics
(Tjoa & Krishnan, 2003)
• Color images (HSI)
• Inspired by the Texture Spectrum method
• Feature vectors formed by 1st order statistics on the
NTU distribution in the I-channel:
• Energy & Entropy
• Mean, Standard deviation, Skew & Kurtosis
• In addition color features C from each color channel C
L2
 C   Hist C (i)
i  L1
L 1
 Hist
i 0
C
(i )
Color Wavelet Covariance
(Karkanis et al, 2003)
• Color images (I1I2I3)
• Discrete Wavelet Frame Transform (DWFT)
on each channel C
• Co-occurrence statistics F on each wavelet band B(k)
• Feature vectors formed by the Covariance of the
cooccurrence statistics between the color channels
B j (k )
Cl ,Cm
CWC

 Cov F
B j (k )
Cl
B j (k )
Cm
,F

Experimental Framework
• We focus only on the textural tissue patterns
• Gastroscopic video 320240 pixels
• Region of interest 128128 pixels
Experimental Framework
• 1,000 Representative video frames
• Verified polyp and normal samples
• 4,000 non-overlapping sub-images 3232 pixels
Experimental Framework
• Support Vector Machines (SVM)
• 10-fold cross validation
• Receiver Operating Characteristics (ROC)
• Accuracy assessed using
the Area Under Characteristic (AUC)
Results
Method
Accuracy (AUC)
1 TS
75.2  2.6 %
2 LBP
80.6  2.5 %
3 TSCHS
87.5  2.1 %
4 CWC
88.6  2.3 %
Results
1
0.9
0.8
Sens itivity
0.7
0.6
0.5
0.4
TS
0.3
LB P
0.2
T S CH S
0.1
CW C
0
0
0.1
0.2
0.3
0.4
0.5
0.6
1 - S p e c ific it y
0.7
0. 8
0.9
1
Conclusions
• We have considered texture as a primary
discriminative feature of gastric polyps
• Four texture feature extraction methods were
considered
• Their performance was compared using SVMs
and ROC analysis
Conclusions
• The development of a CBMS for gastric polyp
detection is feasible
• Color information enhances gastric polyp
discrimination
• The discrimination performance of the spatial
and
the wavelet domain color texture features is
comparable
• The CBMSs developed for colon polyp detection
Thank you