162 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 21, NO.

1, JANUARY 2017

Content-Adaptive Region-Based Color Texture

Descriptors for Medical Images
Farhan Riaz, Ali Hassan, Rida Nisar, Mario Dinis-Ribeiro, and Miguel Tavares Coimbra, Member, IEEE

Abstract—The design of computer-assisted decision

(CAD) systems for different biomedical imaging scenarios
is a challenging task in computer vision. Sometimes, this
challenge can be attributed to the image acquisition mech-
anisms since the lack of control on the cameras can cre-
ate different visualizations of the same imaging site un-
der different rotation, scaling, and illumination parameters,
with a requirement to get a consistent diagnosis by the
CAD systems. Moreover, the images acquired from differ-
ent sites have specific colors, making the use of standard
color spaces highly redundant. In this paper, we propose to
tackle these issues by introducing novel region-based tex-
ture, and color descriptors. The proposed texture features
are based on the usage of analytic Gabor filters (for compen-
sation of illumination variations) followed by the calculation
of first- and second-order statistics of the filter responses
and making them invariant using some trivial mathemati-
cal operators. The proposed color features are obtained by
compensating for the illumination variations in the images
using homomorphic filtering followed by a bag-of-words ap-
proach to obtain the most typical colors in the images. The
proposed features are used for the identification of cancer
in images from two distinct imaging modalities, i.e., gas-
troenterology and dermoscopy. Experiments demonstrate
that the proposed descriptors compares favorably to sev-
eral other state-of-the-art methods, elucidating on the ef-
fectiveness of adapted features for image characterization.
Index Terms—Feature extraction, Gabor filters, Hilbert
transform, pattern recognition, texture analysis.
I. INTRODUCTION
HE extraction of visual characteristics from an image is a
T vital step for computer-assisted diagnostics. Various med-
ical imaging modalities demand the extraction of specific visual
Manuscript received June 30, 2015; revised September 3, 2015;
accepted October 11, 2015. Date of publication October 19, 2015;
date of current version January 31, 2017. This work was supported
by the National University of Sciences and Technology, Islamabad,
Pakistan and Portuguese grants from “Project BRIDGE, Ref CMUP-
EPB/TIC/0069/2013”, “Project Rheumus (Projeto QREN no: 38505)” and
“Project GEMINI, in the scope of PEst-OE/EEI/LA0008/2013 and R&D
Unit 50008 (UID/EEA/50008/2013)”.
F. Riaz, A. Hassan, and R. Nisar are with the Department of
Computer Engineering, College of Electrical and Mechanical Engi-
neering, National University of Sciences and Technology, Rawalpindi
46000, Pakistan (e-mail: farhan.riaz@ceme.nust.edu.pk; alihassan@
ceme.nust.edu.pk; rida@ceme.nust.edu.pk).
M. Coimbra is with the Instituto de Telecomunicações, Department of
Computer Science, Faculdade de Ciłncias da Universidade do Porto,
4169-007, Porto, Portugal (e-mail: mcoimbra@dcc.fc.up.pt).
M. Dinis-Ribeiro is with the CINTESIS/Faculdade de Medicina da Uni-
versidade do Porto and the Instituto Portugal de Oncologia, 4200-319,
Porto, Portugal (e-mail: mario@med.up.pt).
Digital Object Identifier 10.1109/JBHI.2015.2492464
Fig. 1. Dynamic imaging conditions in endoscopy. (a) Image captured
during live exam. (b) Image captured by bringing endoscopic probe
closer to the tissue wall (scale and homogeneous illumination). (c) Arti-
ficially rotated image (adapted from [3]).
features such as shapes and sizes of the anatomical structures
appearing in the images. It is widely known that there are two
fundamental visual descriptors that have been investigated for a
wide range of imaging scenarios: texture and color. The usage
of color and texture information for image description is com-
plementary as it can give better results in various applications
related to computer vision than using only the texture or color
descriptors [1].
The extraction of texture features from medical images is
challenging given the imaging dynamics that involve the varia-
tions appearing in the images due to uncontrollable distance and
angle between the imaging device and the site that is under ob-
servation (see Fig. 1). Scenarios may include (but are not limited
to) gastroenterology [2]–[4], histopathology [5], [6], etc. This
results in various visualizations of a particular image under dif-
ferent rotation, scale, illumination, and illumination gradients
stimulating the need for texture descriptors that are invariant to
such transformations [3].
Additionally, the images acquired from various medical imag-
ing modalities suffer from reduced color spaces [7], [8] (see
Fig. 3). This is because different body structures have their own
typical colors. Therefore, only a small subset of colors in the
whole color space may be available in the medical images (see
Fig. 2). Under these conditions, the standard texture and color
descriptors that are not designed to cater for these imaging dy-
namics are not expected to perform well. The objective in this
paper is to devise the texture and color descriptors that have
the ability to handle the aforementioned imaging dynamics.
More specifically, we have focused on the design of rotation,
scale and illumination invariant texture descriptors based on
Gabor filters and color descriptors that adapt to the underlying
color content of the images, based on a bag-of-words approach.
For demonstrating the effectiveness of the designed visual

2168-2194 © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
RIAZ et al.: CONTENT-ADAPTIVE REGION-BASED COLOR TEXTURE DESCRIPTORS FOR MEDICAL IMAGES
Fig. 2. Image from distinct imaging modalities showing very specific
colors, needing adapted color characteristics for color description (CH -
Chromoendoscopy; NBI - Narrow-band gastroenterology Image).
Fig. 3. Visual illustration of the problem of reduced color spaces. The
RGB cube plot for the images shows that the multimedia images cover
broader parts in the color spaces. The medical images cover limited parts
of the color spaces stimulating the need for adaptive color descriptors.
descriptors (texture and color), we have supported our work
with empirical evidence by considering two different classifi-
cation problems pertaining to distinct imaging modalities: 1)
diagnosis of stomach cancer in vital stained magnification en-
doscopy images, and 2) detection of melanoma in dermoscopy
images.
A. Related Work on Texture Descriptors
The texture features have been extracted using a number of
ways in the past [9]–[11]. The gray level distributions and their
cooccurrences have been used quite often [12], [13] since the
intensities and their variations using the first- and second-order
statistics can yield a good characterization of textures for a wide
range of imaging scenarios. Another type of popular feature
extraction approaches is based on the filter bank approaches,
e.g., Gaussian [14], [15], wavelets [16], etc., that have been
used to highlight specific features in the images.
Recently, different texture descriptors have been proposed in
the literature, some of the popular examples include the scale-
invariant feature transform (SIFT) [17]–[19], the local binary
patterns (LBP) [20]–[22] (and its extensions, e.g., [23], [24],
etc.), the histogram of oriented gradients [25], etc. Such features
have been widely used for various multimedia applications and
although they perform well, there is still room for improvement
163
if the texture descriptors are intended for the description of
medical images. This is because of the medical images have
organic textures exhibiting a high degree of randomness making
them undeterministic. Additionally, a very typical scenario is
the visualization of the images from arbitrary angles and at
varying distances between the tissues and the camera. Therefore,
there is an inherent requirement to design the descriptors, which
are rotation, scale, and illumination (both homogeneous and
nonhomogeneous) invariant.
Traditionally, these requirements are easily fulfilled by filter-
based methods, which exhibit a high degree of diversity to obtain
a directional and multiresolution analysis of the images thus
facilitating the design of rotation and scale invariant texture
descriptors [3], [26] giving good results for the classification of
gastroenterology images. However, if the descriptor is further
adapted to the changing illumination conditions, it is expected
to perform even better.
B. Related Work on Color Descriptors
The research on the design of specific color descriptors is
somewhat limited as compared to that on the texture descriptors.
Earlier approaches to the extraction of color features include the
calculation of histograms of color channels from different color
spaces such as RGB, HSV, HSI, etc. These methods rigidly
quantize the underlying color spaces into a fixed number of bins
and have been widely used in the past for many vision-related ap-
plications [7], [27]–[29]. To perform feature extraction for vari-
ous multimedia applications, such features perform really well.
Lately, however, the researchers have realized that there is the
major drawback in these traditional color descriptors, i.e., they
do not take the nature of color distribution into consideration as
the quantization of color spaces is always fixed (see Fig. 3). The
remedy for this major shortcoming was addressed by the famous
MPEG-7 standard in which the dominant color descriptor was
designed to cater for the changing distribution of color content
in the images [30]. Later on, many adaptive color descriptors
have been designed based on the approaches such as vector
quantization for tailoring the histogram bins according to the
underlying distribution of color contents in the images.
More recently, an applied area that has attracted the attention
of the researchers for the design of adaptive color descriptor is
medical imaging. This is because such images suffer severely
from reduced color spaces thus stimulating the research on more
adequate color descriptors. There have been quite a few exam-
ples of such descriptors that yield much better results as com-
pared to the standard color descriptors that are more adequate
as compared to those designed for the colorful imaging scenes
[7], [8].
The histogram-based color feature extraction methods have
inherent advantages toward rotation and scale invariant feature
extraction as the histograms exhibit these two specific char-
acteristics [31]. Given this particular advantage, most of the
extensions in color features for rotation and scale invariant de-
scription are based on color histograms. For extracting adapted
color features, some authors have resorted to clustering-based
approaches on the standard HSV space to yield color his-
togram bins that are representative of the color distribution in
the images. These methods have been used in several imaging
164 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 21, NO. 1, JANUARY 2017
scenarios such as gastroenterology [3], [7], dermoscopy [32],
[33] just to name a few. Though good results have been obtained
in the literature, one important challenge is to incorporate the
intraimage illumination variations in the image descriptor that
can lead to more suitable color features for the description of
medical images.
C. Contributions
Our main focus in this paper is to design the generic image
descriptors, which encapsulate the texture and color information
in the images. The texture and color descriptors should be able
to handle the rotation, scale, and illumination changes in the
images. Additionally, the color descriptor should adapt itself
to represent the most typical colors that occur in the images.
Correspondingly, the main contributions of this paper are as
follows.
1) A novel rotation, scale, and illumination invariant tex-
ture descriptor is proposed as an extension of a dis-
crete Fourier transform (DFT)-based Gabor descriptor
in [34]. The proposed methodology introduces features
that are robust to complex illumination changes, un-
like [34] where the descriptor design does not effectively
handle nonuniform illumination changes in the images.
These changes are handled by the use of analytic Gabor
filters [35] with a significantly improve low frequency
response that encapsulates the illumination variations in
textures.
2) CIELUV color space separates image color and illumi-
nation components. This fact is exploited to extract illu-
mination invariant color features by prefiltering the “L”
channel using “homomorphic filtering.” The illumination
corrected color image is subjected to vector quantization
(VQ) to identify the most typical colors (“words”) that
appear in the images. The pixels in the novel images are
quantized to these “words” followed by the generation of
adapted color histograms constructed using bag-of-words
approach.
3) The proposed features are used in a classification frame-
work to identify the lesions in images from two distinct
imaging modalities, namely, gastroenterology and der-
moscopy.
The rest of this paper is organized as follows: We discuss the
texture and color features used in this paper for feature extraction
(see Section II) followed by a description of the datasets used
in this paper with our results (see Section III) and conclude the
paper (see Section IV).
II. METHODOLOGIES
Texture and color are the two fundamental visual features that
are of primary importance in the analysis of visual content. In
this section, we will discuss the texture and color features that
we have used for feature extraction from the images.
A. Texture Features
For texture feature extraction, we chose Gabor filters due to
various reasons.
1) First, 2-D Gabor filters are very similar with a simple cell
in the visual cortex of mammals [36]. These cells have
two important properties: their directional selectivity and
bandpass nature. Therefore, they respond to specific spa-
tial frequencies in some specific directions. Mathematical
modeling of these cells lead to bandpass filter bank struc-
tures that are tuned for various orientations and scales
just like Gabor filter banks.
2) Another important characteristic of Gabor filters is that
they have the capability to achieve maximum localiza-
tion in both the space and frequency (this characteristic
of Gabor filters is distinct and not shared by any other
method [37]).
1) Gabor Filters: Gabor filters are obtained as a result of
multiplication of a Gaussian function with a complex Sinusoid.
A 2-D Gabor function g(x, y) and its Fourier transform G(u, v)
can be written as
     
1 1 x2 y2
g(x, y) = exp − + + 2πjW x
2πσx σy 2 σx2 σy2
(1)
  
1 (u − W )2 v2
G(u, v) = exp + (2)
2 σu2 σv2
where σu = 2π1σ x , σv = 2π1σ y , and W is a constant representing
the center frequency of the filter having the highest frequency.
Equation (1) is the product of a Gaussian function with a com-
plex sinusoid. In the frequency domain, we obtain a bandpass
filter where the characteristics of the filter are controlled by the
standard deviation of the Gaussian function and the frequency
of the sinusoid. A Gabor filter bank is formed by obtaining
several filters with varying parameters. The Gabor filter banks
are nonorthogonal due to the nature of the Gaussian function.
Therefore, the redundancy incured by the Gabor filters has to be
reduced to ensure that the information represented by various
filter banks is as distinct as possible. We have used Manjunath’s
[38] strategy to achieve this: design filter parameters such that
the half peak amplitudes of the filters touch each other. This
strategy ensures that the filters capture maximum information
with minimum redundancy. An input image, ξ(x, y) when fil-
tered by the set of Gabor filters g(x, y) is given as

R(x, y) = ξ(x, y)g(x − x1 , y − y1 )dx1 dy1 . (3)
Gabor filters have been widely used for the extraction of
texture features from images [39], [40].
2) Analytic Gabor Filters: The foundation of using the an-
alytic Gabor filters for feature extraction was laid by G. Haley
et al. [35]. Their use is motivated by the fact that the standard
Gabor filters exhibit a potentially significant response at very
low frequencies. This response is undesirable as it exists due to
the inter and intraimage variations in contrast and image inten-
sities. These variations are attributed to the imaging conditions
and not to the texture of the images resulting in misclassification.
There could be several reasons including image sampling with
different average intensity (variation in homogeneous illumina-
tion), illumination gradients, etc. There are two different ap-
proaches to avoid these problems: preprocessing the images, or
RIAZ et al.: CONTENT-ADAPTIVE REGION-BASED COLOR TEXTURE DESCRIPTORS FOR MEDICAL IMAGES 165

 (S )
modifying the Gabor filter banks. Using the former methodol- represented by Rμ (S ) that is a shifted representation of Rμ . It
ogy, the interimage variations can be corrected but the intraim- should be noted that the application of a shift invariant transform

age variations cannot be handled. For these variations, other (S )
on Rμ can give us rotation invariant features for the image. It
image processing methods such as modeling the illumination in is well known that the magnitude of the DFT is shift invariant
the logarithmic domain are needed. [42], and thus, the effect of rotation on the feature vector can be
An effective and more straightforward approach to handle curtailed using DFT of Rμ
(S )
this issue is to modify the Gabor filter to be analytic. This
is accomplished by replacing the real part of g(x, y) with the N
k
inverse Hilbert transform of the imaginary part −ĝim (x, y). FμS R = Rμ(S ) . exp(−j2π n) (9)
n =1
N
gA (x, y) = −ĝim (x, y) + jgim (x, y). (4)
where FμS R is invariant to rotation and scale changes in the
The Fourier transforms of real and imaginary parts of g(x, y) images. Let us now define
are, respectively, conjugate symmetric and conjugate antisym- N
metric resulting in the cancellation for ω ≤ 0.
Rμ(R ) = Rμ k n (10)

G(ω) − G∗ (ω), for ω > 0 n =1
GA (ω) = (5) (R )
0, for ω ≤ 0. where Rμ represents the summation of the aggregated filter
(R )
responses at each scale. Note that Rμ is a vector that is rota-
GA (ω) is the analytic Gabor filter impulse response and has
tion invariant as the responses across all orientations have been
several advantages over the standard Gabor filters for various
summed up. However, if an image is scaled by a certain amount,
applications. 
(R )
1) Improved low frequency response: G(x, y) < GA (ω) at this will be represented by Rμ , which is a shifted representa-
(R )
low frequencies and GA (ω) = 0 at ω = 0. tion of Rμ . The magnitude of the DFT is shift invariant, and
2) Reduced computations as the negative frequency spec- thus, the effect of scale changes on the feature vector can be
(R )
trum does not exist for analytic form of Gabor filters. curtailed using the DFT of Rμ .
All these advantages are acquired without any additional pro- M
cessing. For feature extraction, the images are filtered using the k
FμR S = Rμ(R ) . exp(−j2π n) (11)
analytic form of Gabor filters. M
k =1
r(x, y) = ξ(x, y) ∗ gA (x, y) (6) where FμR S is invariant to rotation and scale changes in the
where r is the response at (x, y), ξ(x, y) is the image, and images. Similarly, the assumption of homogeneousness can be
gA (x, y) is the analytic Gabor filter. extended to calculate the variances of the Gabor filter responses
3) DFT-Based Invariant Features: Assuming that the lo-  2 2 2 2 2
Rσ k n = σ11 σ12 · · · σ1N σ21 · · · σM N (12)
cal texture in an image is spatially homogeneous [41], the Gabor
filter responses for an image region can aggregated by taking which represents a collection of variances of Gabor filter re-
the mean of Gabor filter responses across various scales and sponses in the form of a vector. M is the total number of scales
orientations giving us and N is the total number of orientations for which Gabor filter
 responses have been calculated. As in the case of Rμ , the vec-
Rμ k n = μ11 μ12 · · · μ1N μ21 · · · μM N (7) tor representing the variances for a rotated or scale image will
which represents a collection of means of Gabor filter responses be a shifted representation of Rσ k n . Scale and rotation invari-
in the form of a vector. M is the total number of scales and N is ant texture features based on variances (FσR S and FσS R ) similar
the total number of orientations for which Gabor filter responses to those based on means can be constructed. Concatenation of
have been calculated. For an image that is transformed (rotated rotation and scale invariant features gives us
or scaled), its responses will be represented by a filter that is 
IGabor = FμS R FμR S FσS R FσR S (13)
transformed by the same amount as that of the image. These
transformations are thus indicated by shifts within the feature where IGabor are the novel features that are invariant to rotation,
vector (for rotation or scale changes) indicated in (7). Rotation scale, and illumination changes in the images. The proposed de-
and scale invariant features can be constructed if these shifts are scriptor is an extension of the texture descriptor that has been
effectively catered for in the feature vectors. Let us now define proposed in [34], where only first-order moments of standard
M
Gabor filter responses are used for the extraction of invariant
Rμ(S ) = Rμ k n (8) features. In contrast, this paper proposed the use of analytic
k =1
Gabor filters for the extraction of features that are invariant to
illumination variations, in addition to scale and rotation invari-
where Rμ (S ) represents the summation of the aggregated filter ance. Additionally, the first- and second-order moments are both
(S )
responses at the each orientation. Rμ is a vector, which is scale used for image description. The analytic Gabor filters exhibit
invariant as the responses across all scales have been summed improved low frequency response making the resulting features
up. However, if an image is rotated by a certain angle, this will be very robust to illumination variations in the images.
166 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 21, NO. 1, JANUARY 2017
B. Color Features
For the extraction of color features in the images, we have
used the CIELUV color space. This selection provides two ad-
vantages.
1) Distance between two colors in the CIELUV color space
are perceptually uniform. This can help in the construc-
tion of decision boundaries by the pattern recognition
methodologies, which are more consistent with the hu-
man visual system.
2) In the CIELUV color space, the “L” component repre-
sents the luminosity and the “UV” components represent
the color (chrominance) of a pixel. Due to this decompo-
sition, it is possible to cope with illumination gradients
in the images more easily.
For the medical imaging scenarios, there are two distinct chal-
lenges: 1) a tissue analyzed under varying illumination condi-
tions can create very different visualizations that can potentially
lead to misclassification, and 2) the color features are highly
redundant as most of the organs have very typical colors and
using the full color spaces for image description is not useful.
A better color description of the images can be obtained if the
color features are indicative of the specific color characteristics
of the images. To address these issues, we perform the prepro-
cessing of illumination component of the images followed by
the extraction of specific color features.
1) Preprocessing: The luminosity component in the
CIELUV space can be used to remove illumination changes
in the images. The illumination gradient which is present in the
images is typically composed of low-frequency components.
These components are removed from the images using homo-
morphic filtering [43]. It is based on modeling the image in-
tensity l(x, y) as a combination of the illumination α(x, y) and
reflectance r(x, y) components such that
l(x, y) ∝ α(x, y).r(x, y). (14)
This multiplicative model can be converted to additive model
by taking the logarithmic transformation on both sides
log(l(x, y)) = log(α(x, y)) + log(r(x, y)). (15)
These additive components are considered separable in the
frequency domain. The log(α(x, y)) component is composed
of low-frequency components (illumination gradient), whereas
the log(r(x, y)) is composed of high-frequency components.
An illumination invariant luminance component can thus be
obtained by high pass filtering of log(l(x, y)) followed by an
inverse logarithmic transformation of log(r(x, y)).
2) Vecor Quantization: As discussed previously, uniform
quantization of the color space is not expected to work well
given the very specific colors available in the medical images.
Therefore, we have resorted to the extraction of adapted color
histograms for color feature extraction from the images using
a bag-of-words approach. This approach involves the gener-
ation of codebook using the Linde-Buzo Gray (LBG) vector
quantization algorithm. Given that we have a set of input vec-
tors W = {w1 , w2 . . . wN } ∈ Rk of N distinct points in a k-
dimensional space Rk , a codebook C = {c1 , c2 . . . cM } of M
Algorithm 1 Visual Dictionary
1: procedure CODEBOOK
2: Input: training images
3: for each training image do
4: Convert image to LUV space
5: Perform homomorphic filtering on ‘L’
component
6: Initialize codebook C to data mean, Th and 
7: p ← 1, where p is number of partitions
8: while E{||W − C(W )||2 > Th do
9: Replace μj as {μj }pj=1 +  and {μj }pj=1 − 
10: Set p ← 2p
11: end while
12: M ← p2 words
13: end for
14: end procedure
code words is obtained giving us a spatial visual vocabulary.
To construct C, the LBG algorithm starts with one code vector
(average of the training data) and iteratively bisects the code
vector in power of 2. The distortion monotonically decreases in
the LBG algorithm at each iteration. An important consideration
is the use of an appropriate distortion measure and we use the
mean squared error between the sample and the initial clusters
as our measure of distortion, i.e.,
d = arg minE{||W − C||2 } (16)
C(W )
where W are the data points and C is the distance of each point
from each word in the codebook.
3) Histogram of Codebook: Once the codebook is gener-
ated, every image for which the features have to be calculated is
first preprocessed using conversion to the CIELUV color space
followed by homomorphic filtering of the luminance compo-
nent. Later, every pixel in the image is compared with all the
words available in the codebook as follows:
D = [d1 , d2 , . . . dN ] (17)
where each di is an M -dimensional vector indicating the dis-
tance of the ith pixel from the mth word in the codebook. A
histogram of these code words is later generated that is the un-
derlying color feature that is used for the extraction of adapted
color features from the images.
C. Classification
In our experiments, we have used supervised classification
methods for the characterization of tissues. Supervised methods
involve two steps: a training phase in which some part of the data
are used to characterize the underlying model that it follows and
a testing phase in which the model is applied on unseen data to
see how well does the model fit the novel data. Many approaches
to supervised classification exist and their choice depends on
the underlying applications. In this study, we have evaluated
the performance of four different classifiers for our application,
namely, the one nearest neighbor (1NN), naive Bayes (NB),
RIAZ et al.: CONTENT-ADAPTIVE REGION-BASED COLOR TEXTURE DESCRIPTORS FOR MEDICAL IMAGES 167

Algorithm 2 Bag of Words

1: procedure ADAPTED HISTOGRAM
2: Input: Images for histogram computation
3: for each image do
4: Convert image to LUV space
5: Perform homomorphic filtering on “L”
component
6: for n ← 1 to N data points do
7: pick nth data point wn
8: calculate dn ← (wn − C)2
9: quantize wn to codeword with minimum d
10: end for
11: W ← {w1 , w2 . . . wN }
12: generate histogram of quantized W Fig. 4. Dinis-Ribeiro proposal for the classification of chromoen-
13: end for doscopy.
14: end procedure

support vector machine (SVM) with a linear kernel (parameter

optimization using cross validation), and decision trees (DT).

III. EXPERIMENTS
In this paper, we have used two datasets from very differ-
ent imaging scenarios: vital stained magnification endoscopy
for the identification of gastric cancer and dermoscopy for the
classification of melanoma. Among texture features, we have
used 2DDFT [26], AHT [44], Gabsum [34], LBP [45], and bag-
of-words-based SIFT [46] descriptors for comparison purposes.
It is a complementary set of region-based and local visual de-
scriptors. For color feature extraction, we have used standard
LUV (LUV), adapted LUV (ALUV), and adapted illumination
invariant (ALUVill) histograms. Due to the limited data that we
have from the two imaging modalities (specifically the cases
of cancer), we have used cross validation for our experiments.
The statistical relevance of our experiments was calculated by
repeating the cross validation experiments 50 times and per-
forming a T-test for statistical significance (threshold to 0.05).
Classification results are obtained using four difference clas-
sifiers to ensure that obtained results are not classifier model
specific.
A. Vital Stained Magnification Endoscopy
One of the potential applications of the proposed descriptor
is its usage for feature extraction from gastroenterology images.
This is a complicated problem given that various organs hav-
ing distinct structures result in various perspectives of viewing
the same tissue at various scales from various angles. Addition-
ally, various imaging modalities focus on complementary visual
characteristics of the images that are useful for diagnosis. Given
the dynamics in the GE scenario due to the organs and imaging
modalities, we have used images from vital stained magnifi-
cation endoscopy (Chromoendoscopy—CH) for diagnosis of
stomach cancer. In CH, the full visible spectrum of the light is
used for tissue illumination and tissue enhancement is achieved
using dyes (e.g., methylene blue) for improving the visibility of
Fig. 5. Vital stained magnification endoscopy images along with their
manual annotations. The images suffer severe illumination problems
that are partially solved by the manual annotations. Even within the
annotations, there is uneven illumination in the images.
the gastric mucosa. The CH images are clinically classified as
normal (Group I), precancerous (Group II) or cancerous (Group
III) based on the color and texture of the images according to
Dinis–Ribeiro classification proposal for CH images [47] (see
Fig. 4). The CH images were acquired by the physicians at the
Portuguese Institute of Oncology (IPO) Porto, Portugal, during
routine clinical work (Please refer to [3] for more details on
the data). The CH dataset consists of 130 images, which are
distributed as: 31% (40) Group I, 57% (75) Group II, and 12%
(15) Group III images.
For our classification experiments, we have used only the
manually annotated image regions (see Fig. 5). Feature extrac-
tion was performed on full images, followed by the selection
of features from those pixels that lie in the regions representing
the clinical annotations for the respective images. The objective
of the classification task is to identify the cancerous tissues,
i.e., classify the images as either normal (Group A) or poten-
tially cancerous (Group B). Consequently, the images in the DR
proposal (see Fig. 4) belonging to Group I should be classi-
fied as Normal (Group A), whereas those belonging to Group
II or III should be classified as potentially cancerous (Group
B). The weka data mining tool was used for the classification
results presented in this paper. All the results were obtained
using tenfold cross validation. We have used the overall classi-
fication accuracy and area under the receiver operating charac-
teristics (ROC) curves for evaluation of the performance of the
168 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 21, NO. 1, JANUARY 2017

TABLE I
OVERALL CLASSIFICATION RESULTS OBTAINED USING THE PROPOSED DESCRIPTOR FOR GASTROENTEROLOGY IMAGES

1NN NB SVM DT

Features TP Rate ROC Area TP Rate ROC area TP Rate ROC Area TP Rate ROC Area

IGabor 0.89* 0.87* 0.90* 0.93* 0.90* 0.95* 0.90* 0.88*

2DDFT 0.89 0.87 0.85* 0.92* 0.90* 0.94* 0.89* 0.88
AHT 0.89* 0.86* 0.87* 0.92* 0.90* 0.94* 0.89* 0.88*
Gabsum 0.87* 0.85* 0.76* 0.91* 0.88* 0.91* 0.87* 0.85*
SIFT 0.67* 0.74* 0.72* 0.89* 0.77* 0.79* 0.72* 0.76*
LBP 0.84* 0.79* 0.77* 0.85* 0.90* 0.92* 0.83* 0.83*
ALUVill 0.90 0.87* 0.82* 0.93* 0.90* 0.93* 0.89* 0.86*
ALUV 0.89 0.85* 0.83* 0.89* 0.87* 0.92* 0.87* 0.83*
LUV 0.78* 0.74* 0.80* 0.89* 0.78* 0.89* 0.78* 0.74*
IGabor+ALUVill 0.90 0.88 0.92 0.94 0.93 0.96 0.91 0.89

The statistical significance of all texture and color features are calculated against IGabor+ALUVill. All statistically significant
results are marked with “* .”

classifier. However, it should be noted that a fair comparison can
be done using the area under ROC curve given that the dataset
has unequal instance of normal and cancer/precancer lesions.
The statistical significance of the results is also calculated using
the T-test.
When only texture features are used for feature extraction,
the IGabor features outperform all the other methods that have
been considered in this paper (Table I). We owe this perfor-
mance to the illumination invariant nature of IGabor features
that is done by using analytic Gabor filters that improve the low
frequency response of the filters. Even when using the simple
classifier such as 1NN and NB, the IGabor features show good
performance elucidating on the fact that the features are linearly
separable and not model specific. The best performance is ob-
tained when IGabor features are used in combination with an
SVM classifier. However, it should be noted that the difference
in the performance of IGabor from other texture features is not
very large. This is because we use only the manually annotated
image regions for feature extraction and this partly solves the
problem of complex illumination changes in the images (see
Fig. 5). Additionally, the 2DDFT, AHT, and Gabsum features
are invariant to homogeneous illumination variations in the im-
ages already but illumination variations within the images are
not effectively handled by these methods in contrast to that of
IGabor. It should be noted that the IGabor that is an extension of
Gabsum (use of analytic filters improving low frequency re-
sponses and using standard deviation of filter responses in ad-
dition to means) demonstrates significantly good performance
using any classifier showing the superiority of IGabor over
Gabsum.
When color features are used for classification, LUV features
show poor performance in the classification of lesions (Table I).
When ALUV features are used a significant performance im-
provement is obtained, whereas when illumination variations are
handled, a further performance improvement is obtained. Our
results are consistent with the nature of the vital stained magni-
fication endoscopy images where the images have very specific
color features that essentially encompass a very small range of
colors in the color space. The observation is consistent across all
the classifiers elucidating on the fact that the proposed features
are not specific to any classification model. Given that the IGabor
and ALUVill show the best performance as texture and color
descriptors, respectively, we use them in combination to assess
if they represent complementary image characteristics for the
classification. Our experiments show that the best performance
is obtained when a combination of texture and color features
are used for feature extraction. The results are always better and
statistically significant than using only texture or color features.
An important observation is a relatively inferior performance
of local descriptors (LBP and SIFT) indicating toward the lack
of discriminative information between the normal and cancer
tissues at higher frequencies (microtexture) owing to the texture
noise/distortion in the images [48]–[50]. Our observation was
complemented by conducting experiments in which the features
obtained only from lower and medium frequency bands (macro-
texture) in the Gabor filter were used as image descriptors (for
Gabsum, AHT, and 2DDFT) giving similar classification re-
sults. This validated our suspicion that most of the descriptive
content for cancer lies in lower frequency bands (macrotexture).
B. Dermoscopy
Another medical imaging modality that shares the imaging
dynamics of gastroenterology is Dermoscopy. Dermoscopy is a
clinical procedure that is usually used to identify the presence
of melanoma in patients. Melanoma is considered as one of the
deadliest forms of skin cancer as an increase in its incidence
and the consequent mortality has become one of the major con-
cerns for public health. Researchers have shown that the two
visual descriptors, which are imperative in the identification of
melanoma are texture and color (details in [51]). The most vital
imaging dynamics to be catered for by the visual descriptors
for this imaging scenario are the rotation, scale, and illumi-
nation invariance [32]. The need for rotation invariance arises
from the fact that an imaging site can be visualized from arbi-
trary angles and this variation should not manifest itself in the
resulting features. Additionally, different images of the same le-
sion acquired from the same equipment and even with the same
zooming factor can represent considerably “different” lesions to
the viewer as it is difficult to guarantee consistent illumination as
even a mild pressure from the dermoscope can cause significant
RIAZ et al.: CONTENT-ADAPTIVE REGION-BASED COLOR TEXTURE DESCRIPTORS FOR MEDICAL IMAGES 169

TABLE II
OVERALL CLASSIFICATION RESULTS OBTAINED USING THE PROPOSED DESCRIPTOR FOR DERMOSCOPY IMAGES

1NN NB SVM DT

Features TP Rate ROC Area TP Rate ROC Area TP Rate ROC Area TP Rate ROC Area

IGabor 0.81* 0.70* 0.85 0.89* 0.87* 0.91* 0.82* 0.71*

2DDFT 0.79* 0.67* 0.84* 0.87* 0.85* 0.87* 0.82* 0.70*
AHT 0.81* 0.70* 0.83* 0.87* 0.86* 0.88* 0.81* 0.68*
Gabsum 0.75* 0.60* 0.83 0.85* 0.85* 0.87* 0.83* 0.79
SIFT 0.52* 0.55* 0.50* 0.65* 0.66* 0.60* 0.52* 0.58*
LBP 0.69* 0.50* 0.80* 0.58* 0.80* 0.63* 0.80* 0.62*
ALUVill 0.85* 0.76* 0.80* 0.85* 0.86* 0.92* 0.85* 0.79*
ALUV 0.84* 0.74* 0.80* 0.84* 0.87* 0.92* 0.86* 0.80*
LUV 0.78* 0.64* 0.78* 0.83* 0.83* 0.87* 0.77* 0.70*
IGabor+ALUVill 0.86 0.80 0.86 0.90 0.88 0.93 0.88 0.81

The statistical significance of all texture and color features are calculated against IGabor+ALUVill. All statistically significant
results are marked with “* .”

IGabor texture features show best classification results among

Fig. 6. Ground truth for dermoscopy: original image representing
melanoma (left) and manually segmented mask of image region indi-
cating melanoma (right).
deformation of the skin [33]. Irrespective of these changes, we
want to get the same (or near similar) features for a particu-
lar lesion. The color characteristics of the typical nevus and
melanoma significantly differ, and the colors are very specific
(encompassing a very small area in the color space) making
the proposed color descriptor interesting for application to the
dermoscopy images. In this paper, our specific objective is to
classify a skin lesion as being either a nevus or melanoma.
The dataset that we have used is composed of 200 dermoscopy
images with the following composition: 80% (160) nevus and
20% (40) melanoma. The images were acquired at the Hospi-
tal Pedro Hispano, Matosinhos [52], [53]. All images have been
acquired during clinical exams using a dermoscope at a magnifi-
cation of 20x with a resolution of 765 × 573 pixels. Each image
was manually segmented (to identify the lesions) and classified
by an experienced dermatologist as being normal, atypical ne-
vus (benign) or melanoma (malignant). For our classification
experiments, we have used only the manually annotated image
regions (see Fig. 6—The methodology is similar as described for
CH images). The objective of the classification task is to identify
the presence of melanoma in the patients. The weka data mining
tool was used for the classification results presented in this paper.
All the results were obtained using tenfold cross validation. We
have used the overall classification accuracy and area under the
ROC curves for evaluation of the performance of the classifier.
The latter provides a more fair assessment of the results given
that dataset has unequal instances of normal and melanoma.
For dermoscopic image classification, our observations from
the results are consistent with that of GE images (Table II). The
all the texture features. We owe this performance to the robust-
ness of the texture features to the illumination variations in the
images. The superiority of IGabor is consistent across all the
classifiers showing that the results are not specific to any classi-
fier model. Among color features, the adapted features (ALUV
and ALUVill) always show better performance as compared to
the standard color descriptors (LUV) by a significant margin.
When a combination of texture and color features is used, the
best performance is obtained irrespective of the classifier used
and the results are always significantly different. When local
descriptors (LBP and SIFT) are used for skin lesion character-
ization, relatively lower classification accuracies are observed.
This is because dermoscopy images are composed of some ar-
tifacts such as hair, skin lines, etc., which are emphasized on
by these descriptors, whereas they do not contribute to lesion
characterization giving us inferior performance. We repeated
the experiments performed for CH images on the dermoscopy
dataset in which only the features from lower and medium fre-
quency bands were used for classification purposes. Our findings
for dermoscopy are consistent with those of CH images.
IV. CONCLUSION
This paper deals with feature extraction from the images in
those imaging modalities, which exhibit a high degree of imag-
ing dynamics. These dynamics can potentially arise in those
scenarios where maintaining full control of the camera is typ-
ically not possible, or various visualization of the same tissue
is possible because there is no sense of strict direction in the
image contents. For such applications, the image descriptors
should be invariant to the rotation, scaling, and illumination
changes in the images. This has led us to the design of invariant
texture descriptor, the IGabor, which is based on Gabor fil-
ters. For the compensation of inter- and intraimage illumination
variation, the analytic Gabor filters have been used. For rota-
tion and scale invariance, the homogeneous texture features are
calculated followed by summation of coefficients across scale
and taking magnitude of DFT across orientation, and summa-
tion of coefficients across orientation followed by magnitude of
DFT across scale. The resulting texture features are invariant to
170 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 21, NO. 1, JANUARY 2017
illumination, scale, and rotation changes in the images. For color
feature extraction, the images are transformed to the LUV color
space to decouple luminance and color. The illumination varia-
tions are compensated using homomorphic filtering on the “L”
channel followed by the calculation of adapted color features.
The features are calculated by using some “representative” im-
ages to calculate their most typical color characteristics using
LBG algorithm to generate a codebook. Pixels in the remaining
images are quantized to the words available in the codebook
followed by the generation of a histogram of codebooks that is
later used for classification of the images.
The proposed feature set is used for classifying images from
two distinct imaging modalities, carrying very different objec-
tives but sharing the imaging dynamics: chromoendoscopy and
dermoscopy. Our experiments have demonstrated the superiority
of both texture and color features over other methods considered
in this paper. We have also used a combination of both texture
and color features and shown that the proposed features outper-
form the other methods that have been considered in this paper,
elucidating on the generic nature of the proposed descriptors.
In addition to rotation and scale invariance, the descriptors that
compensate for illumination variations in the images tend to
work better. Last but not the least, our experiments demonstrate
that the proposed features are not specific to a classification
model as they almost always perform better than other feature
extraction methods irrespective of the classifier used.
ACKNOWLEDGMENT
The authors would like to thank Portuguese Institute of On-
cology, Porto, Portugal, for their support during the manual
annotation of the dataset.
REFERENCES
[1] J. Y. Choi, Y. M. Ro, and K. N. Plataniotis, “Color local texture features
for color face recognition,” IEEE Trans. Image Process., vol. 21, no. 3,
pp. 1366–1380, Mar. 2012.
[2] Y. Shen, P. Guturu, and B. P. Buckles, “Wireless capsule endoscopy video
segmentation using an unsupervised learning approach based on proba-
bilistic latent semantic analysis with scale invariant features,” IEEE Trans.
Inform. Technol. Biomed., vol. 16, no. 1, pp. 98–105, Jan. 2012.
[3] F. Riaz, F. Silva, M. Ribeiro, and M. Coimbra, “Invariant Gabor texture
descriptors for classification of gastroenterology images,” IEEE Trans.
Biomed. Eng., vol. 59, no. 10, pp. 2893–2904, Oct. 2012.
[4] S. Atasoy, D. Mateus, A. Meining, G.-Z. Yang, and N. Navab, “Endoscopic
video manifolds for targeted optical biopsy,” IEEE Trans. Med. Imag.,
vol. 31, no. 3, pp. 637–653, Mar. 2012.
[5] S. H. Raza, R. M. Parry, R. A. Moffitt, A. N. Young, and M. D. Wang,
“An analysis of scale and rotation invariance in the bag-of-features method
for histopathological image classification,” in Medical Image Computing
and Computer-Assisted Intervention. Berlin, Germany: Springer, 2011,
pp. 66–74.
[6] J. C. Caicedo, A. Cruz, and F. A. Gonzalez, “Histopathology image clas-
sification using bag of features and kernel functions,” in Artificial Intelli-
gence in Medicine. New York, NY, USA: Springer, 2009, pp. 126–135.
[7] A. Sousa, M. Dinis-Ribeiro, M. Areia, and M. Coimbra, “Identifying can-
cer regions in vital-stained magnification endoscopy images using adapted
color histograms,” in Proc. IEEE 16th Int. Conf. Image Process., Nov.
2009, pp. 681–684.
[8] B. Zhang, B. Kumar, and D. Zhang, “Detecting diabetes mellitus and
nonproliferative diabetic retinopathy using tongue color, texture, and
geometry features,” IEEE Trans. Biomed. Eng., vol. 61, no. 2, pp. 491–501,
Feb. 2014.
[9] M.-C. Yang, W. K. Moon, Y.-C. Wang, M. S. Bae, C.-S. Huang, J.-H. Chen,
and R.-F. Chang, “Robust texture analysis using multi-resolution gray-
scale invariant features for breast sonographic tumor diagnosis,” IEEE
Trans. Med. Imag., vol. 32, no. 12, pp. 2262–2273, Dec. 2013.
[10] H. Kong, M. Gurcan, and K. Belkacem-Boussaid, “Partitioning
histopathological images: An integrated framework for supervised color-
texture segmentation and cell splitting,” IEEE Trans. Med. Imag., vol. 30,
no. 9, pp. 1661–1677, Sep. 2011.
[11] M. Sadeghi, T. Lee, D. Mclean, H. Lui, and M. Atkins, “Detection and
analysis of irregular streaks in dermoscopic images of skin lesions,” IEEE
Trans. Med. Imag., vol. 32, no. 5, pp. 849–861, May 2013.
[12] Z. HongQing, “Segmentation of blood vessels in retinal images using 2D
entropies of gray level-gradient cooccurrence matrix,” in Proc. IEEE Int.
Conf. Acoust. Speech Signal Process., vol. 3, May 2004, pp. iii-509–iii-
512.
[13] S. Park, B. Kim, J. Lee, J. M. Goo, and Y.-G. Shin, “Ggo nodule volume-
preserving nonrigid lung registration using GLCM texture analysis,” IEEE
Trans. Biomed. Eng., vol. 58, no. 10, pp. 2885–2894, Oct. 2011.
[14] F. van der Lijn, M. de Bruijne, S. Klein, T. den Heijer, Y. Hoogendam,
A. van der Lugt, M. M. B. Breteler, and W. Niessen, “Automated brain
structure segmentation based on atlas registration and appearance models,”
IEEE Trans. Med. Imag., vol. 31, no. 2, pp. 276–286, Feb. 2012.
[15] R. Bansal, L. Staib, D. Xu, H. Zhu, and B. Peterson, “Statistical analyses
of brain surfaces using Gaussian random fields on 2-D manifolds,” IEEE
Trans. Med. Imag., vol. 26, no. 1, pp. 46–57, Jan. 2007.
[16] S. Dua, U. Acharya, P. Chowriappa, and S. Sree, “Wavelet-based en-
ergy features for glaucomatous image classification,” IEEE Trans. Inform.
Technol. Biomed., vol. 16, no. 1, pp. 80–87, Jan. 2012.
[17] D. Ai, X. Han, X. Ruan, and Y. wei Chen, “Adaptive color independent
components based sift descriptors for image classification,” in Proc. 20th
Int. Conf. Pattern Recog., Aug. 2010, pp. 2436–2439.
[18] T. Takahashi, T. Kawano, K. Ito, T. Aoki, and S. Kondo, “Performance
evaluation of a geometric correction method for multi-projector display
using sift and phase-only correlation,” in Proc. 17th IEEE Int. Conf. Image
Process., Sep. 2010, pp. 1189–1192.
[19] W. Lu and M. Wu, “Multimedia forensic hash based on visual words,” in
Proc. 17th IEEE Int. Conf. Image Process., Sep. 2010, pp. 989–992.
[20] A. Sousa, “Analysis of color and texture features of vital stained magnifi-
cation endoscopy images for computer assisted diagnosis of precancerous
and cancer lesions,” Ph.D. Thesis, Faculdade de Engenharia da Universi-
dade do Porto, Porto, Portugal, 2008.
[21] T. Ojala, M. Pietikainen, and D. Harwood, “A comparative study of tex-
ture measures with classification based on feature distributions,” Pattern
Recog., vol. 29, no. 1, pp. 51–59, Jan. 1996.
[22] F. Riaz, M.-D. Ribeiro, P. Pimentel-Nunes, and M. Tavares Coimbra,
“Integral scale histogram local binary patterns for classification of narrow-
band gastroenterology images,” in Proc. IEEE 35th Annu. Int. Conf. Eng.
Med. Biol. Soc., Jul. 2013, pp. 3714–3717.
[23] W.-H. Liao, “Region description using extended local ternary patterns,”
in Proc. IEEE 20th Int. Conf., 2010, pp. 1003–1006.
[24] L. Nanni, A. Lumini, and S. Brahnam, “Local binary patterns variants
as texture descriptors for medical image analysis,” Artif. Intell. Med.,
vol. 49, no. 2, pp. 117–125, 2010.
[25] N. Dalal and B. Triggs, “Histograms of oriented gradients for human
detection,” in Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recog.,
vol. 1, Jun. 2005, pp. 886–893.
[26] F. Riaz, A. Hassan, S. Rehman, and U. Qamar, “Texture classification
using rotation- and scale-invariant Gabor texture features,” IEEE Signal
Process. Lett., vol. 20, no. 6, pp. 607–610, 2013.
[27] S. Jayaram, S. Schmugge, M. Shin, and L. V. Tsap, “Effect of colorspace
transformation, the illuminance component, and color modeling on skin
detection,” in Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recog.,
vol. 2, Jun. 2004, pp. II-813–II-818.
[28] A. Ferman, A. Tekalp, and R. Mehrotra, “Robust color histogram descrip-
tors for video segment retrieval and identification,” IEEE Trans. Image
Process., vol. 11, no. 5, pp. 497–508, May 2002.
[29] W. Wang and M. Hua, “Extracting dominant textures in real time with
multi-scale hue-saturation-intensity histograms,” Image IEEE Trans. Im-
age Process., vol. 22, no. 11, pp. 4237–4248, Nov. 2013.
[30] B. Manjunath, J.-R. Ohm, V. Vasudevan, and A. Yamada, “Color and
texture descriptors,” IEEE Trans. Circuits Syst. Video Technol., vol. 11,
no. 6, pp. 703–715, Jun. 2001.
[31] C.-H. Yao and S.-Y. Chen, “Retrieval of translated, rotated and scaled
color textures,” Pattern Recog., vol. 36, no. 4, pp. 913–929, 2003.
RIAZ et al.: CONTENT-ADAPTIVE REGION-BASED COLOR TEXTURE DESCRIPTORS FOR MEDICAL IMAGES
[32] H. Zhou, M. Chen, and J. Rehg, “Dermoscopic interest point detector and
descriptor,” in Proc. IEEE Int. Symp. Biomed. Imag., Nano Macro, 2009,
pp. 1318–1321.
[33] F. Peruch, F. Bogo, M. Bonazza, V.-M. Cappelleri, and E. Peserico,
“Simpler, faster, more accurate melanocytic lesion segmentation through
MEDS,” IEEE Trans. Biomed. Eng., vol. 61, no. 2, pp. 557–565, Feb.
2014.
[34] F. Riaz, M. Dinis-Ribeiro, P. Pimentel-Nunes, and M. T. Coimbra, “A
DFT based rotation and scale invariant Gabor texture descriptor and its
application to gastroenterology,” in Proc. Int. Conf. Image Process., 2013,
pp. 1443–1446.
[35] G. Haley and B. Manjunath, “Rotation-invariant texture classification us-
ing a complete space-frequency model,” IEEE Trans. Image Process., vol.
8, no. 2, pp. 255–269, Feb. 1999.
[36] D. J. Field, “Relations between the statistics of natural images and the
response properties of cortical cells,” IEE J. Radio Commun. Eng., vol. 4,
no. 12, 1987.
[37] J. G. Daugman, “High confidence visual recognition of persons by a test
of statistical independence,” IEEE Trans. Pattern Anal. Mach. Intell., vol.
15, no. 11, pp. 1148–1161, Nov. 1993.
[38] P. Wu, B. S. Manjunath, S. Newsam, and H. D. Shin, “A texture descriptor
for browsing and similarity retrieval,” J. Signal Process.: Image Commun.,
vol. 16, no. 1, pp. 33–43, 2000.
[39] R. Porter and N. Canagarajah, “Robust rotation-invariant texture classifi-
cation: Wavelet, Gabor filter and GMRF based schemes,” IEE Proc. Vis.
Image Signal Process., vol. 144, no. 3, pp. 180–188, Jun. 1997.
[40] J. K. Kmrinen, “Feature extraction using Gabor filters,” Ph.D. disser-
tation, Lappeenranta University of Technology, Lappeenranta, Finland,
2003.
[41] B. S. Manjunath and W. Y. Ma, “Texture features for browsing and retrieval
of image data,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 18, no. 8,
pp. 837–842, Aug. 1996.
[42] A. V. Oppenheim, R. W. Schafer, and J. R. Buck, Discrete-Time Signal
Processing. Englewood Cliffs, NJ, USA: Prentice-Hall, 1999.
[43] T. Schuchert, T. Aach, and H. Scharr, “Range flow in varying illumination:
Algorithms and comparisons,” IEEE Trans. Pattern Anal. Mach. Intell.,
vol. 32, no. 9, pp. 1646–1658, Sep. 2010.
171
[44] F. Riaz, “Assisted analysis of gastroenterology images using computer
vision methodologies,” Ph.D. dissertation, Faculdade de Ciencias da Uni-
versidade do Porto, Porto, Portugal, 2012.
[45] T. Ojala, M. Pietikäinen, and T. Mäenpää, “Multiresolution gray-scale and
rotation invariant texture classification with local binary patterns,” IEEE
Trans. Pattern Anal. Mach. Intell., vol. 24, no. 7, pp. 971–987, Jul. 2002.
[46] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, and A. Vedaldi, “Describ-
ing textures in the wild,” in Proc. IEEE Conf. Comput. Vis. Pattern Recog.,
2014, pp. 3606–3613.
[47] M. D. Ribeiro, “Clinical, endoscopic and laboratorial assessment of pa-
tients with associated lesions to gastric adenocarcinoma,” Ph.D. disserta-
tion, Faculdade de Medicina da Universidade do Porto, Porto, Portugal,
2005.
[48] G. Bao, Y. Ye, U. Khan, X. Zheng, and K. Pahlavan, “Modeling of the
movement of the endoscopy capsule inside GI tract based on the captured
endoscopic images,” presented at the Int. Conf. Modeling, Simulation,
Visualization Methods, Las Vegas, Las Vegas, NV, USA, 2012.
[49] D. Koppel, Y.-F. Wang, and H. Lee, “Image-based rendering and model-
ing in video-endoscopy,” in Proc. IEEE Int. Symp. Biomed. Imag, Nano
Macro, 2004, pp. 269–272.
[50] A. Vécsei, G. Amann, S. Hegenbart, M. Liedlgruber, and A. Uhl, “Auto-
mated marsh-like classification of celiac disease in children using local
texture operators,” Comput. Biol. Med., vol. 41, no. 6, pp. 313–325, 2011.
[51] I. Maglogiannis and C. N. Doukas, “Overview of advanced computer
vision systems for skin lesions characterization,” IEEE Trans. Inf. Technol.
Biomed., vol. 13, no. 5, pp. 721–733, Sep. 2009.
[52] C. Barata, M. Ruela, M. Francisco, T. Mendonca, and J. Marques, “Two
systems for the detection of melanomas in dermoscopy images using
texture and color features,” IEEE Syst. J., vol. 8, no. 3, pp. 965–979, Sep.
2013.
[53] C. Barata, J. Marques, and J. Rozeira, “A system for the detection of
pigment network in dermoscopy images using directional filters,” IEEE
Trans. Biomed. Eng., vol. 59, no. 10, pp. 2744–2754, Oct. 2012.
Authors’ photographs and biographies not available at the time of
publication.