Open Journal of Me di cal Imaging, 2011, 1, 26-42
doi:10.4236/ojmi.2011.12005 Published Online December 2011 (http://www.SciRP.org/journal/ojmi)
Copyright © 2011 SciRes. OJMI
Textural Based SVM for MS Lesion Segmentation in
FLAIR MRIs
Bassem A. Abdullah1, Akmal A. Younis1, Pradip M. Pattany2, Efrat Saraf-Lavi2
1Department of Electrical and Computer Engineering, University of Miami, Miami, USA
2Department of Radiology, Miller School of Medicine, University of Miami, Miami, USA
E-mail: b.abdullah@umiami.edu
Received October 17, 2011; revised November 8, 2011; accepted December 3, 2011
Abstract
In this paper, a new technique is proposed for automatic segmentation of multiple sclerosis (MS) lesions
from brain magnetic resonance imaging (MRI). The technique uses textural features to describe the blocks of
each MRI slice along with position and neighborhood features. A trained support vector machine (SVM) is
used to discriminate between the blocks in regions of MS lesions and the blocks in non-MS lesion regions
based on mainly the textural features with aid of the other features. The MRI slice blocks’ classification is
used to provide an initial segmentation. A comprehensive post processing module is then utilized to refine
and improve the quality of the initial segmentation. The main contribution of the proposed technique de-
scribed in this paper is the use of textural features to detect MS lesions in a fully automated process without
the need to manually define regions of interest (ROIs). In addition, the post processing module is generic
enough to be applied to the results of any other MS segmentation technique to improve the segmentation
quality. This technique is evaluated using ten real MRI data-sets with 10% used in the training of the tex-
tural-based SVM. The average results for the performance evaluation of the presented technique were 0.79
for dice similarity, 0.68 for sensitivity and 0.9 for the percentage of the detected lesion load. These results
indicate that the proposed method would be useful in clinical practice for the detection of MS lesions from
MRI.
Keywords: MRI, Texture Analysis, Segmentation, Multiple Sclerosis, SVM, ROI, Post-Processing, Fuzzy
Rules
1. Introduction
Multiple sclerosis (MS) is a chronic idiopathic disease
that results in multiple areas of inflammatory demyeliza-
tion in the central nervou s system (CNS) [1]. Progressive
MS lesion formation often leads to cogn itive decline and
physical disability. Due to its sen sitivity in detecting MS
lesions, Magnetic Resonance Imaging (MRI) has become
an effective tool for diagnosing MS and monitoring its
progression [2,3]. Accurate manual assessment of each
lesion in MR images would be a demanding and labori-
ous task, and would also be subjective and have poor
reproducibility [4]. Automatic Segmentation offers an
attractive alternative to manual segmentation which re-
mains a time-consuming task and suffers from intra- and
inter-expert variability [5]. However, the progression of
the MS lesions shows considerable variability and MS
lesions present temporal changes in shape, location, and
area between patients and even for the same patient [6-9].
This makes the automatic segmentation of MS lesions a
challengin g problem.
Because of the importance of computer-aided MS le-
sion detection, a number of semi-automated and auto-
mated methods have been proposed for segmenting MS
lesions in MR images [4,5,10-16]. Texture analysis in
MRI has been used with some success in neuro-imaging
to detect lesions and abnormalities. Textural analysis
refers to a set of processes applied to characterize special
variation patterns of voxels grayscale in an image. Tex-
tural features have been used [17] to differentiate be-
tween lesion white matter (LWM), normal white matter
(NWM) and normal appearing white matter (NAWM).
Texture classification was also used for the analysis of
multiple sclerosis [1]. The use of textural features is
promising to provide good results in MS lesion segmen-
tation, especially if the Fluid Attenuated Inversion Re-
B. A. ABDULLAH ET AL.27
covery (FLAIR) sequence is used where the textural at-
tributes of the MS lesions are different from other re-
gions in the brain.
However, to the best of our knowledge, texture based
approaches that have been previously reported were ap-
plied to regions of interest (ROIs) that are manually se-
lected by an expert to indicate potential regions including
MS lesions, which makes the segmentation process
semi-au tomated. As such, efforts are needed to automate
the use of textural features in the detection of MS le-
sions.
In this paper, we propose a technique that uses textural
features to describe the blocks of each MRI slice along
with position and neighborhood features. A trained clas-
sifier is used to discriminate between the blocks and de-
tect the blocks that potentially in clude MS lesio n s mainly
based on the textural features with aid of the other fea-
tures. The blocks classification is used to provide an ini-
tial coarse segmentation of the MRI slices. The textural-
based classifier is built using Support Vector Machine
(SVM), one of the widely used supervised learning algo-
rithms that have been utilized successfully in many ap-
plications [18,19]. Classification errors can occur in this
initial coarse segmentation. False positives can arise be-
cause of the similarity in textural attributes between
some healthy regions in the brain and the MS lesions
used in training the textural-based classifier. On contrary,
false negatives can arise from any source of noise in the
MR images that may corrupt the textural attributes of the
MS lesions. To overcome both types of errors, a com-
prehensive post processing module is added to improve
the quality of the initial segmentation by addressing each
type of error individually to generate the final MS lesion
segmentation. All the sequences of MRI are used in the
task of MS segmentation. FLAIR images especially axial
slices are selected in our system. Due to the higher accu-
racy of the FLAIR imaging sequence in revealing MS
lesions and assessing the total lesion load [20,21], axial
FLAIR MRI were used in this paper.
The paper consists of five sections including this in-
troduction section. In Section 2, the details of the pro-
posed segmentation technique are illustrated. The ex-
perimental results are presented in Section 3 and dis-
cussed in Section 4. The paper conclusion is stated in
Section 5. For completeness, Appendix A provides the
details for calculating the textural features.
2. Materials and Methods
2.1. Dataset
The dataset used in this paper involves FLAIR MRI se-
quences for ten subjects (4 males, age range: 50 - 72 and
6 females, age range: 30 - 59). On average, each FLAIR
MRI sequence consists of thirty seven slices that cover
the whole brain. All the subjects were referred for brain
MRI studies based on an earlier diagnosis of MS. The
MS lesions were manually labeled by a neuroradiologist.
The ten MRI studies were acquired using a 3.0T MR
scanner under a human subject’s protocol approved by
the institutional review board of th e Univ ersity Of Miami
Miller School Of Medicine (Florida, USA). The axial
FLAIR sequences used in this paper were acquired using
the following imaging parameters: 9000/103/2500/256 ×
204/17/123 (repetition time ms/echo time ms/inversion
time/matrix size/echo time length/imaging frequency). In
clinical practice, the T2-weighted and/or FLAIR images
have been an established routine sequence for diagnosis
of MS lesions [22], and thus FLAIR MR imaging was
employed in this paper. All images were acquired with a
slice thickness of 3 mm, an interslice gap of 0.9 mm, and
a field of view of 175 × 220 mm. Pixels are sampled to
16-bits and resolution of 408 × 512 (pixel size 0.43 ×
0.43 mm).
2.2. MS Lesions Segmentation Framework
The proposed MS lesions segmentation framework is
described in Figure 1. The MRI FLAIR slices of the
brain are preprocessed for intensity correction to remove
the effect of noise and differences in brightness and con-
trast between different scans of different subjects. The
next step is the main processing module which is used
for the detection of initial MS lesions regions based on
textural features. This stage generates scores for each
voxel in the slices that repr esents the prob ability of being
MS voxel or not. The connected voxels having non-zero
scores form regions of MS lesions. The post processing
step involves addressing false positiv e MS regions based
on location attributes and detecting false negative MS
regions through inter-slice comparisons using the 3D
nature of the MRI. After that, the post processing step
corrects the MS lesions for each slice by removing vox-
els based on distance and grayscale features, and adding
neighboring voxels using region growing based on gray-
scale features and adding voxels by removing holes in
lesion regions using lesion continuity fact.
2.3. Preprocessing
Due to different operating conditions from a subject to
another, brightness and contrast of slices may vary
among subjects. This affects performance of segmenta-
tion that is based on textural features which are calcu-
lated based on grayscale intensities. If a dataset is used
for training, better histogram matching of the dataset to
Copyright © 2011 SciRes. OJMI
B. A. ABDULLAH ET AL.
Copyright © 2011 SciRes. OJMI
28
on improving a slice from MS6 (subject dataset to be
segmented) with reference to MS3 (subject dataset used
in training).
In addition, the preprocessing step includes the detec-
tion of the center of mass and the sagittal plane (central
line passing through the center in the same direction of
the slice orientation) for each slice. These geometrical
parameters are needed in feature extraction.
2.4. Textural Based Detection of Initial MS
Lesions Regions
Each preprocessed MRI slice, is processed by a trained
detector to get initial MS lesions regions. The detector
engine in our method is implemented using support vec-
tor machine. Training the detector machine is done by
processing the training dataset and dividing its slices into
square blocks and assigning binary class for each block.
If the block contains at least one pixel labeled as MS, it
is classified as MS block (class 1). Otherwise, the block
is classified as non-MS block (class 0) if all of its pixels
are labeled as non-MS pixels. Each block is described by
a feature vector which mainly represents textural features
of the block. During segmentation, the slice to be seg-
mented is divided into square overlapping blocks and
each block is classified by the trained engine as MS
block or non-MS block.
Figure 1. MS lesion segmentation framework.
2.4.1. Block Size
be segmented and the training set will lead to error re-
duction. Statistics were previously made to measure the size of
the multiple sclerosis lesions. The co mmon v a lues for the
diameter are between 3.5 mm and 13.5 mm [24]. For any
input MRI studies, the square w × w blocks are selected
automatically to tightly cover the smallest possible di-
ameter as a function of the given pixel sizes calculated
from the input resolution and field of view. For the size
We used our preprocessing technique used before in
[23] that starts with applying contrast-brightness correc-
tion to maximize the intersection between the histogram
of the training and segmentation datasets followed by
using 3D anisotropic filter to avoid empty histogram bins.
Figure 2 shows the effect of the preprocessing technique
(a) (b) (c)
Figure 2. Preprocessing: (a) A slice from the reference subject MS3 (used in training); (b) A slice from subject MS6 before
preprocessing; (c) The same slice of MS6 after preprocessing.
B. A. ABDULLAH ET AL.
Copyright © 2011 SciRes. OJMI
29
of pixels in our studies, square 8 × 8 (pixel2) blocks (3.4
× 3.4 mm2) tightly cover the smallest possible MS lesion
diameter.
2.4.2. Fea tur e V ector
To describe each square block of the FLAIR MRI slice, a
feature vector of thirty four features is calculated. The
block features are categorized into three categories: twenty
four textural features, two position features and eight
neighboring blocks features. The features are summa-
rized in Table 1. The textural features include histo-
gram-based features (mean and Variance), gradient-based
features (gradient mean and gradient Variance), run
length-based features (gray level non-uniformity, run
length non-uniformity) and co-occurrence matrix-based
features (contrast, entropy and absolute value). Run leng-
th-based features are calculated 4 times for horizontal,
vertical, 45 degrees and 135 degrees directions. Co-oc-
currence matrix-based features are calculated using a
pixel distance d = 1 and for the same angles as the run
length-based features. The details for calculating the
textural-based features are provided in Appendix A.
The position features are the slice relative location
with reference to the bottom slice and the radial Euclid-
ean distance between the block top pixel and the center
of the slice normalized by dividing it by the longest di-
ameter of the slice. A sample labeled slice to illustrate
the position features is shown in Figure 3. In Figure
3(a), the block is labeled by 2, the slice center is labeled
by 1, the radial Euclid ean distan ce is lab eled b y 3 and the
longest diameter is labeled by 4. The center and the
longest diameter of the slice are parameter geometrically
calculated in the preprocessing step.
Table 1. Feature vector.
Features
Category Features
Textural features 1 Mean (Histogram based feature)
2 Variance (Histogram based feature)
3 Gradient mean (Gradient b as e d fe a t u re)
4 Gradient variance (Gradient based feature)
5 - 8 Grey level non-uniformality
in the 4 directions (Run length based feature)
9 - 12 Run length non-uni for mality
in the 4 directions (Run length based feature)
13 - 16 Contrast in the 4 directions
(Co-occurrence matrix based featu re)
17 - 20 Absolute value in the 4 dire cti ons
(Co-occurrence matrix based featu re)
21 - 24 Entropy in the 4 directions
(Co-occurrence matrix based featu re)
Position features 25 Slice relative location
26 Normalized radial distance
between block and slice center
Neighboring
blocks features
27 - 34 Differences in grayscale
between the block and
each of the 8 neighboring blocks
The neighboring blocks features are the difference
between the mean grayscale of the current block and the
mean grayscale of each of the eight neighboring blocks.
In Figure 3(b), the 3 × 3 grid circled by yellow circle
(a)
(b)
Figure 3. Position and neighboring blocks features extr-
action on sample slices. (a) Calculation of the normalized
radial distance betw ee n block (2) and slice center (1) (length
of line 3/length of line 4); (b) Calculation of the eight neigh-
boring blocks features (difference between mean grayscale
of the centered red block and mean grayscale of each of the
eight neighboring green blocks in the grid circled by the
yellow circle).
B. A. ABDULLAH ET AL.
30
demonstrates the current bl ock colored wi th red color and
its eight neighboring blocks colored with green color.
2.4.3. SVM Tr ai ni ng and Segmentation
Support Vector Machine (SVM) is a supervised learning
algorithm, which has at its core a method for creating a
predictor function from a set of training data where the
function itself can be a binary, a multi-category, or even
a general regression predictor. To accomplish this mathe-
matical feat, SVMs find a hypersurface which attempts
to split the positiv e and negative examples with the larg-
est possible margin on all sides of the hyperplane. It uses
a kernel function to transform data from input space into
a high dimensional feature space in which it searches for
a separating hyperplane. The radial basis function (RBF)
kernel is selected to be the kernel of the SVM. This ker-
nel nonlinearly maps samples into a higher dimensional
space so it can handle the case when the relation between
class labels and attributes is nonlinear. The library libsvm
2.9 [25] include all the methods needed to do the imple-
mentation, training and prediction tasks of the SVM. It is
incorporated in our method to handle all the SVM opera-
tions.
2.4.3. 1. Tra i ning
The dataset of one or more subjects is used to generate
the SVM training set. The slices of this training dataset
are divided into n square blocks of size w × w pixels.
SVM Training set T is composed of training entries ti (xi,
yi) where xi is the feature vector of the block bi, yi is the
class label of this block for i =1: n (number of blocks in-
cluded in the training set). Our classification problem is
binary, so yi is either 0 or 1. The training entry is said to
be positive entry if yi is 1 and negative in the other case.
For each slice of the training dataset, each group of con-
nected pixels labeled manually as MS pixels forms a
lesion region .
Blocks involved in the positive training entries (TP)
are generated by localizing all the lesion regions and for
each of them, the smallest rectangle that encloses the
lesion region is divided into non-overlapping square
blocks of size w × w pixels. Each block bi of these blocks
is labeled by yi = 1 if any of the w2 pixels inside this
block is manually labeled as MS pixel. Any block that
contains at least 1 MS pixel is defined in our method as
MS block.
Similarly, the blocks involved in the negative training
entries (TN) are generated by localizing the non-back-
ground pixels that are not manually labeled as MS pixels
and dividing them into non-overlapping square blocks of
size w × w pixels. Each block bi of these blocks is la-
beled by yi = 0. Feature vector xi is calculated for each
block of both positive and negative training entries. The
positive training entries Tp contain blocks that contain 1:
w2 MS pixels. This helps the SVM engine to learn the
features of the blocks that either partially or completely
contain MS pixels.
In our case, we used one subject dataset (only 10% of
the subjects) which consists of thirty seven slices as the
training dataset. Since the training set entries are as many
as the number of blocks, the training set will be large
enough (134173 training entries against 34 features with
ratio 3946:1) to avo id the curse of dimensionality, which
is the problem that the performances of the pattern clas-
sification systems could deteriorate if the ratio of the
number of training data to that of features used for the
classifier is relatively small [26].
The training set entries were fed to the SVM engine to
generate a MS classifier which is able to classify any
square w × w block of a brain FLAIR MRI slice as MS
block (y = 1) or non-MS block (y = 0) based on its fea-
ture vector (x).
2.4.3. 2. Seg mentation
Each of the slices of the datasets to be segmented is di-
vided into overlapping square blocks of size w × w pixels.
The feature vector for each block is calculated. The
trained SVM is used to predic t the class labels for all the
overlapping blocks. The block division is done in an
overlapping manner to detect any possible MS blocks.
For any block classified as MS block, assuming true pos-
itive classification, this does not mean that all pixels of
the block should be classified as MS pixels because the
SVM engine is trained to detect the blocks that contains
MS pixels completely o r partially. For each slice, all pix-
els are assigned an intege r score. This score is initialized
with a zero value. During segmentation, if any block is
classified as MS block (y = 1), the scores of all pixels
inside the block are incremented. As the blocks are
overlapped, each pixel is part of w2 blocks as demon-
strated in Figure 4. Thus, the score will be any value
from 0 to w2.
After classification, these scores act as initial lesion
probability maps where a large score indicates high
probability for the pixel to be an MS pixel. Initial MS
lesions can be generated by assigning any pixel of non-
zero score as MS pixel. Figure 5 shows segmentation of
sample slice from subject (MS6). Figure 5(a) shows the
preprocessed FLAIR slice. Figure 5(b) is the ground truth
for the lesions generated through manual segmentation.
Figure 5(c) is the initial segmentation by considering
any pixel of score higher than zero as MS pixel. Figure
5(d) provides a colored evaluation of the segmentation
where the true positive pixels are marked by blue, false
positive are marked by red, false negatives are marked
by green and true negatives are the background pixels.
Copyright © 2011 SciRes. OJMI
B. A. ABDULLAH ET AL.
Copyright © 2011 SciRes. OJMI
31
2.5. Post Processing
The purpose of the post processing step is to improve
and refine the performance of initial segmentation through
dealing with different types of errors (false positives
and false negatives). Figure 6 shows the initial seg-
mentation of a sample slice from subject MS5 (Figure
6(a)) and the colored evaluation of the segmentation in
which the false negatives and positives are marked in
green and red colors, respectively (Figure 6(b)). Errors
in the initial segmentation of MS lesions can be classi-
fied as:
Type 1: False negatives resulting from not detecting
MS lesion regions (labeled by 1 in Figure 6(b)).
Figure 4. All possible overlapping blocks that contain a
pixel: for 8 × 8 blocks (w = 8), the red pixel is part of w2 =
64 blocks. The eight bold blocks are samples where the red
pixel lies in the coordinates (8,8) of block 1, (7,7) of block
2 … and (1,1) of block 8.
Type 2: False negatives resulting from incomplete
MS lesion regions (labeled by 2 in Figure 6(b)).
Type 3: False positives resulting from false MS lesion
regions (labeled by 3 in Figure 6(b)).
(a) (b) (c) (d)
Figure 5. Initial MS lesions regions dete ction: (a) Preproce ssed slice fr om MS6; (b) Ground truth; (c ) Initial segmentation; (d)
Colored evaluation of segmentation.
(a) (b)
Figure 6. False negatives and positives in the textural segmentation (a) initial segmentation of a slice from MS5 and (b)
colored evaluation of the initial segmentation where the different types of errors are labeled by a number matching the
orresponding error type. c
B. A. ABDULLAH ET AL.
Copyright © 2011 SciRes. OJMI
32
Type 4: False positives resulting from false portions
of true MS lesion regions (labeled by 4 in Figure
6(b)).
The following steps, which constitute the post proc-
essing of the MS lesion segmentation framework, are a
set of logical operations that aim to address the different
types of errors in the initial segmentation of MS lesions
without adding new errors. The subject dataset used in
training (MS3) was segmented by the textural SVM to
get the initial segmentation which was analyzed for the
different errors in MS lesions segmentation to formulate
the criteria and thresholds used in post processing. This
is summarized in the block diagram shown in Figure 7.
Step1: Elimination of false positives resulting from
lesion region s i n unc ommon locations
In this step, errors of type 3 in MS lesions are ad-
dressed. This type of errors in MS lesions results from
detected MS lesion regions which are completely false.
Some of these MS lesion regions that are located in un-
common locations can be eliminated. Odd locations in-
clude MS lesions outside the brain area, close to the
brain boundary, or close to the sagittal plan [27]. In Fig-
ure 8, step 1 of post processing is applied to the initial
segmentation of a slice from subject MS6, shown in
Figure 8(a) and color evaluated in Figure 8(b), to elimi-
nate the erroneous MS lesion regions circled in yellow
circle as they are located so close to the boundary of the
slice. The same slice after applying step 1 of post proc-
essing is shown in (Figures 8(c) and (d)).
Step 2: Detection of non-detected MS lesion regions
(false negatives)
In this step, errors of type 1 in MS lesions are ad-
dressed. This type of errors in MS lesions results from
not detecting the lesion region, i.e., completely missing it.
According to [28], in most cases the MS lesions extend
only into one to three consecutive slices when the thick-
ness of the slices is 3 mm. Therefore, In order to recover
the missing MS lesion regions the initial segmentations
of the previous and the next slices (or neighboring slices)
are considered. The detected MS lesion regions in the
previous and next slices are intersected based on the
common coordinates on both slices generating a new
slice of MS lesion regions. Any pixel in the MS lesion
regions resulting from the intersection is assigned the
average of the segmentation scores of the two pixels lo-
cated at the same slice local coordinates in the intersect-
ing slices. Each intersection lesion region is assigned a
score which is the average of the scores of the pixels in
the lesion region. Because of the 3D nature of the MRI
slices and the fact that the lesion occupies a volume, the
lesion regions generated from the intersection should be
highly correlated to the lesion regions in the original
slice (the slice between the intersecting slices) especially
if the generated lesions are of high scores. If this inter-
section leads to new regions in the current slice that have
high scores, there will be a high probability that these
Figure 7. Formulation of criteria and thresholds used in post proce ssing.
B. A. ABDULLAH ET AL.
Copyright © 2011 SciRes. OJMI
33
new lesion regions are part of non-detected lesions in the
current slice, and they should be added as initial seg-
mentations. In Figure 9, step 2 of post processing is ap-
plied to the initial segmentation of a slice from subject
MS6. Figure 9(a) shows slice 12 and the circled green
lesion is a sample for a completely non-detected MS le-
sion region. Figures 9(b) and (c) show slices 11 and 13,
respectively. The circled lesion regions in both of them
are detected lesion regions in the textural segmentation
step. When these lesion regions are intersected as shown
in Figure 9(d), they recover part of the non-detected
lesion region in slice 12 as shown in Figure 9(e). The
recovered part of lesion region can act as a seed that can
be expanded in the region growing used as part of post
processing step 3.
Step 3: Lesion Regions Shape Correction
In this step, errors of type 2 and 4 in MS lesions are
addressed. These two types of errors represent false parts
in the segmented MS lesion regions in th e form of either
additional parts that need to be removed or incomplete
parts that need to be detected. Both types of errors are
addressed through shape correction of each segmented
MS lesion region without adding or deleting MS lesion
regions. Each detected MS lesion region in each slice is
processed to correct its shape through the elimination of
false positive pixels (type 4) and adding non-detected or
false negative pixels (type 2). The shape correction of
detected MS lesions is performed through the following
three operations. Figure 11 will be used to illustrate the
application of the different operations in post processing
step 3 to a sample segmented slice from MS6. Figures
11(i1) and (i2) show the initial colored evaluation of the
segmentation and the initial segmentation, respectively.
The other parts of Figure 11 will be used to illustrate the
corresponding operations in the following discussion of
step 3.
(a) Elimination of the false positives on the bound-
ary of detected MS lesion region s.
The colored evaluation of the segmentation depicted in
Figure 11(i1) shows that each lesion region colored in
blue area (true positiv e) is surrounded by a red boundary
(false positive). These false positive pixels on the bound-
ary of the lesion regions may arise from the similarity
between the textural properties (features) of non-MS
regions and MS lesion regions. These false positive pix-
els may also arise from blocks classified in the initial
(a) (b) (c) (d)
Figure 8. Post processing step 1: elimination of segmented lesion regions in odd locations. (a) and (b) are the slice seg-
mentation and colored evaluation of the segmentation before applying step 1 (MS lesion regions in odd locations are circled in
yellow). (c) and (d) are the slice segmentation and colored evaluation of the segmentation after applying step1.
(a) (b) (c) (d) (e)
Figure 9. Post processing step 2: detection of the non-detected lesion regions using neighboring slice s. (a) Non detected lesion
in slice 12 of MS6. (b) and (c) detected lesions in the slices 11 and 13 respectively are intersected to recover part of the non
detected lesion region. (d) Intersection provides part a new lesion region. (e) Portion of the lesion region in the slice 12 is
ecovered by adding (a) and (d). r
B. A. ABDULLAH ET AL.
Copyright © 2011 SciRes. OJMI
34
segmentation as MS blocks which will cause all the
block pixels’ scores to be incremented increasing their
probability to be MS-pixel even when some of the pixels
in the block are not MS pixels. To eliminate these false
positive pixels, esp ecially on the boundary, the following
parameters are considered for each pixel:
1) Euclidean distance between the pixel and the lesion
region boundary.
2) Difference between the grayscale of the pixel and
the mean grayscale of the lesion region.
3) Segmentation score of the pixel at the conclusion of
the initial segmentation step.
Some of the pixels of the non-MS pixels on the bound-
ary of lesion regions can be eliminated based on these
parameters using a fuzzy engine designed for this pur-
pose where the membership functions and the thresh-old
values are calculated based on the analysis of the initial
segmentation of the training subject dataset (MS3). A
Summary of the fuzzy engine including variable fuzzi-
fication, fuzzy rules and defuzzification is provided in
Figure 10. The fuzzy rules output is the decision which
is binary variable have two values; either keep the pixel
Figure 10. Fuzzy engine used in lesion regions shape correction (step 3a): variables fuzzification, fuzzy rules and defu-
zzification. (In fuzzy rules, X indicates don’t care condition and DECISION = KEEP means keep the pixel in the MS lesion
region and DECISION = REMOVE means remove the pixel from the MS lesion region).
B. A. ABDULLAH ET AL.35
(i1) (a1) (b1) (c1)
(i2) (a2) (b2) (c2)
Figure 11. Post processing step 3: lesion regions shape correction. (i1) and (i2): The initial colored evaluation of the
segmentation and the initial segmentation; (a1) and (a2): Effect of applying step 3a; (b1) and (b2): Effect of applying step 3b;
(c1) and (c2): Effect of applying step 3c.
in the lesion region (KEEP) or remove it (REMOVE).
The defuzzification is performed using the centroid rule
which is used in case of classification [29]. Figures
11(a1) and (a2) show the effect of applying the fuzzy
engine operation on a slice from MS6. The lesion regions
became smoother after trimming the false positive pixels
on the boundary, but some pixels inside the lesion re-
gions were eliminated by mistake leaving some holes
(green voxels in figure 11-a1 and black voxels in Figure
11(a2). These holes will be addressed in operation (c) of
step 3 of post processing, to be discussed later. Any ex-
cessive pixels trimmed from the boundary can be recov-
ered in operation (b) of step 3 of post processing where
false negatives are addressed.
(b) Elimination of the false negatives on the boun-
dary of the lesion regions.
The colored evaluation of the segmentation depicted in
Figure 11(i1) shows green pixels (false negatives) con-
nected to some of the lesion regions blue areas (true
positives). These false negativ e pixels on the boundary of
the lesion regions can arise from the dissimilarity be-
tween the textural properties (features) of the non-de-
tected lesion areas and the textural properties of the de-
tected lesion region itself. To recover these pixels, region
growing is applied for each lesion region. Any of the
pixels neighboring to each lesion region are included in
the closely adjacent MS lesion region if the absolute dif-
ference between the grayscale of the pixel and the mean
grayscale of the lesion region does not exceed the stan-
dard deviation of the grayscale of the lesion region. The
region growing operation is rep eated recursively until no
new pixels are added. Figures 11(b1) and (b2) sh ow the
effect of applying this recursive region growing opera-
tion, where most of the false negative pixels were recov-
ered.
(c) Forcing lesion continuity to eliminat e false nega-
tives (holes) inside the lesion regions.
For each lesion region, there may be some pixels in-
side the region which are not detected in the initial seg-
mentation (due to the block’s textural features being dif-
ferent from features of MS blocks) or detected but re-
moved as part of the elimination of false positives on the
boundary of MS lesions during operation (a) of this step
while trimming the lesion region. By applying the logical
concept of lesion continu ity, which means that the lesion
cannot have inside holes, all the pixels inside the bound-
ary of the lesion regions are assigned to be MS pixels.
Figures 11(c1) and (c2) show the effect of applying this
operation where all the holes were filled.
All the operations visualized in Figure 11 shows that
Copyright © 2011 SciRes. OJMI
B. A. ABDULLAH ET AL.
36
step 3 did not add or remove any lesion regions from the
initial segmentation but only the lesion regions became
more completed and smoother.
2.6. Evaluation of the Proposed Method
To evaluate the performance of the proposed segmenta-
tion method, the dice similarity, sensitivity and percent-
age of the detected lesion load are calculated. The dice
similarity (DS) is a measure of the similarity between the
manual segmentation (X) and the automatic segmenta-
tion (Y). The equation for the calculation can be written
as: DS = 2

X
YXY
As stated in [29,30], a DS score above 0.7 is generally
considered as very good, especially when the segmented
structures are small. Dice Similarity is calculated in our
evaluation module twice. The first is calculated based on
the similarity of voxels (DSV) and the second is calcu-
lated based on the similarity of lesion regions (DSR). DSV
uses the number of common MS voxels between manual
and automatic segmentation for
X
Yand uses the
number of MS voxels of manual and automatic segmen-
tation for
X
and Y respectively. DSR uses the
number of common MS lesion regions between man- ual
and automatic segmentation for
X
Y and uses the
number of MS lesion regions of manual and auto- matic
segmentation for
X
and Y respectively. In the con-
text of DSR, the automatically segmented lesion region
that shares at least one pixel with a manually segmented
lesion region is considered as a common MS lesion re-
gion since the number of MS lesion regions is more
clinically relevant than the number of voxels [4]. For
example, applying these definitions of DSV and DSR to
the initial segmentation s shown in Figures 11(i1) and (i2)
yields values of 0.73 (good segmentation) and 1.0 (per-
fect segmentation), respectively.
Sensitivity is a measure of how many lesions are de-
tected. It can be calculated as the percentage of true posi-
tive voxels to the total number of MS voxels in the
ground truth.
Percentage of detected lesion load is a measure of how
much lesion volume is detected compared to the original
lesion volume. The detected lesion volume takes into
account all the positive lesions whether true or false.
Having a percentage of detected lesion load close to 1.0
is clinically satisfactory since it provides a relatively
accurate measure of the MS lesions volume.
3. Results
The dataset of ten real FLAIR MRI axial sequences were
used to evaluate the performance of the proposed seg-
mentation method. The performance metrics detailed in
Section 2.6 were calculated. The segmentation results are
summarized in Table 2. In this table, for each study su b-
ject, the dice similarity based on lesion regions DSR, the
dice similarity based on voxels DSV, sensitivity and de-
tected lesion load are given. Overall average is given for
each of these performance metrics.
The average metrics are 0.79 for DSR, 0.71 for DSV,
0.68 for sensitivity and 0.9 for percentage of detected
lesion load. The average detected lesion load indicates
that the proposed method could detect the MS lesion
with reasonable error rates. Although the average dice
similarity based on voxels DSV is 0.71, which exceeds the
minimum value for reasonably good segmentation, there
were drops in the performance for some of the studies.
Based on the analysis of the results for these studies, the
MS lesions were found to be very small for these studies
(percentage of MS lesions volume in voxels to the total
volume in voxels less than 0.1%).
Excellent result for the segmentation of the training set
is a bottom line for accepting the technique. If the seg -
mentation result of the training dataset (MS3) is removed
from the average calculation, the average metrics would
be (0.77 for DSR, 0.69 for DSV, 0.66 for sensitivity and
0.88 for percentage of detected lesion load) which is still
a very good result. However, it is included for compare-
son with results of other techniques that includes the
training set segmentation result in their averages.
The effect of the post processing steps on the overall
performance is shown in Figure 12. For each study, the
dice similarity based on Voxels (DSV) is calculated be-
fore and after the use of the post processing step. The
average improvement in dice similarity of the overall
Table 2. Segmentation result.
Study
Dice Similarity
based on
Number of
Lesion Regions
Dice Similarity
based on
Number of
Voxels
Sensitivity
based on
Voxels
Detected
Lesion Load
based on
Voxels (%)
MS20.88 0.78 0.67 0.93
MS30.96 0.93 0.89 1.1
MS40.68 0.64 0.72 0.92
MS50.72 0.68 0.59 0.65
MS60.84 0.72 0.64 0.76
MS70.77 0.67 0.63 0.91
MS80.68 0.68 0.72 1.05
MS90.76 0.63 0.57 0.78
MS100.83 0.71 0.62 0.81
MS110.75 0.68 0.74 1.12
Average0.79 0.71 0.68 0.90
Copyright © 2011 SciRes. OJMI
B. A. ABDULLAH ET AL.
Copyright © 2011 SciRes. OJMI
37
Figure 12. Effect of the post processing steps on the overall segmentation performance.
segmentation due to the post processing stage varies is
12%.
4. Discussion
A novel method for MS lesions segmentation in FLAIR
sequence of brain MR images has been developed. The
segmentation process goes through three steps. The first
step is the preprocessing which is done to improve the
brightness and contrast of all the FLAIR slices of the
subject dataset. The second step is the main processing
which involves using a trained SVM to detect the initial
MS lesions from the individual slices using feature vec-
tor which are mainly composed of textural features. The
third step is the post processing that aims to improve and
refine the performance of the initial segmentation gener-
ated through the SVM-based textural segmentation. In
that regard, the post processing step addresses all possi-
ble types of errors in the MS lesion segmentation results
of the second step in order to reduce the overall errors
including both false positives and false negatives.
The main processing classifier uses thirty four features
in three categories. These categories of features are se-
lected to have analogy with the features used non-inten-
tionally by the expert in the task of manual labeling of
MS areas. According to our observations, when the ex-
pert labels MS lesions in the FLAIR slice, the hyper in-
tense areas are the potential areas to hav e the lesion. This
is emulated in our technique by using the group of twenty
four textural features. Candidate areas are filtered based
on previous experience with the brain positions where
most likely lesions occur; hence a group of two position
based features is used. Besides, the expert takes into ac-
count the difference between the intensity of the lesion
area and the neighboring areas intensities to take final
decision and we emulate this by using the eight neigh-
boring features. Although both textural features and
neighboring features are based on intensity, no redun-
dancy exists between them as the first group is used to
aid the classifier in the detection of special pattern areas
while the later is used to take into account the relation of
the intensity of the area and the neighboring areas.
The main processing classifier uses an SVM engine.
SVM Parameters selection and training set balancing
directly affect the classification performance. The SVM
penalty parameter C and the RBF kernel parameter γ are
chosen via a grid search using cross validation as pro-
posed in [25]. In cross validation, the training dataset are
divided into subsets. Sequentially one subset is tested
using the classifier trained on the remaining subsets.
Cross validation accuracy is the percentage of data which
are correctly classified. Various pairs of (C; γ) valu es are
tried in the cross validation tests and the one with the
best cross validation accuracy is picked. We applied the
cross validation on the training subject (MS3) dataset.
Using this offline exhaustive search, C and γ that pro-
vided best accuracy in the cross validation for our tech-
nique are 1 and 0.029 respectively. On the other hand,
the training set used in training the SVM is highly im-
balanced. The size of the negative training entries (TN)
is much higher than th e positive training entries (TP) due
to the relative size of lesion with respect to the normal
brain tissues. This affects the performance of SVM.
However traditional approaches to overcome imbalanced
data involve either over-sample the minority class (MS
blocks) or under-sample the majority class (non-MS
B. A. ABDULLAH ET AL.
38
blocks). The first results in a distribution that no longer
approximates the target distribution and the later results
in discarding instances that may contain valuable infor-
mation. Our decision was to leave the data with neither
over-sampling nor under-sampling to avoid biasing the
classifier and to keep the real distribution. Future im-
provement should address balancing the dataset with no
added inaccuracy.
Receiver operating characteristics (ROC) analysis was
performed to evaluate the main processing classifier that
generates the initial MS-blocks. Due to using overlap-
ping blocks, each slice pixel is included in 64 blocks (in
case of using square blocks of 8 × 8 pixels). A score is
given to each pixel equals to the number of blocks that
encloses the pixel and classified as MS-block. To draw
the ROC curve, a threshold is defined as the number of
positive b locks needed to co nsider the pixel as MS pixel.
This threshold was changed from 64 blocks down to 0,
and for each case the specificity (true negatives rate) and
sensitivity (true positives rate) were calculated in order
to create the ROC curve as plotted in Figure 13 where
the false positive rate (1-specificity) is on x-axis and the
sensitivity is on y-axis. Using very high threshold leads
to zero false positives and very low sensitivity wh ile very
low threshold leads to both very high false positive rate
and sensitivity. For all tests, the ROC curve falls above
the diagonal indicating good classification.
Our first contribution in this method is using textural
features without manual selection of ROI which was an
area for future research and improvement [1]. To the best
of our knowledge, th e common use of tex tural featur es in
MS lesions detection was previously attempted with aid
of manual selection of regions of interests (ROIs). In the
approach presented in this paper, overlapping square
blocks, with adaptively determined size, are used to re-
place these ROIs. Connected blocks classified as MS
blocks form a lesion region. In the post processing step,
these lesion regions are trimmed to remove extra pixels
and/or extended to include undetected pixels, as appro-
priate, to finally arrive at accurate, smooth, and complete
lesion regions without the need for manual selection of
ROIs.
The second contribution of the presented method is th e
post processing step that is both generic and modular
which means it is independent on the previous steps.
Thus, it can be generalized and applied with any other
MS lesion segmentation technique to improve the per-
formance. In this step, different types of errors in MS
lesion segmentation are addressed based on logical han-
dling of both false positives and false negatives. When
used with any other technique, the score parameter used
in our approach can be replaced by any other parameter
that provides the prob ability of the pixel being part of an
MS lesion based on the other technique.
One of the recent publications [4] provides a compare-
son table between different techniques for detection of
the multiple sclerosis lesions according to the dice simi-
larity based on lesion regions. We quote th e table with in
Table 3 with our results added as the last line. For each
technique, the citation is referenced and the methods
used in segmen tation are prov id ed along with the number
of subjects used in the evaluation and the average dice
similarity obtained using the technique. In the original
publication of [4], the dice similarity was calculated
based on the common regions between the manual seg-
mentation and the automatic segmentation and similarly
Figure 13. ROC (Receiver Operating Characteristics) curve for main processing performance.
Copyright © 2011 SciRes. OJMI
B. A. ABDULLAH ET AL.39
Table 3. Comparison of the automated methods for detection of MS in MR images.
Authors Segmentation Method No. of studiesAverage Dice Similarity (based on lesion regions)
Boudraa et al. 2000 [ 1 3] Fuzzy C-Means 10 0.62
Leemput et al. 2001 [14 ] Stochastic model 50 0.51
Zijdenbos et al. 2002 [15] Pipeline analysis 29 0.68
Khayati et al. 2008 [21] AMM (adaptive mixtures method),
MRF (Markov r andom field model) 20 0.75
Yamamoto et al. 2010 [4] Region growing,
LSM (level set method) 6 0.77
Proposed method Texture Analysis,
SVM (support vector machines) 10 0.79
we also used the value of DSR in the Table 3. The table
shows that our method has an average regional dice
similarity of 0.79 which is the highest among the past
studies. This does not mean that our method is the best in
terms of automatic segmentation performance because
the comparative results in Table 3 are dependent on the
image properties of the datasets, which are different
among the techniqu es includ ed in the tab le. Howeve r, the
comparison shows the success of our method for detect-
ing MS lesions in real MRI datasets with competitive
results.
5. Conclusions
We have developed an automated method for detection
of MS lesions in brain MR images using textural based
SVM. The main contributions of the presented method
are using textural features without manual selection of
ROI and the comprehensive post processing step that
handles different types of errors in MS lesions segmenta-
tion that can be generalized to improve the performance
of any other MS segmentation technique. The method
has been tested using ten real FLAIR MRI datasets. The
performance evaluation and comparative results with
other automated techniques show that our method pro-
vides competitive results for the detection of MS lesions.
6. Acknowledgements
This work was supported in part by National Institutes of
Health (NIH), National Institu te of Neurological Disease
and Stroke (NINDS) through grant #R41NS060473.
7. References
[1] J. Zhang, L. Z. Tong, L. Wang and N. Li, “Texture
Analysis of Multiple Sclerosis: A Comparative Study,”
Magnetic Resonance Imaging, Vol. 26, No. 8, 2008, pp.
1160-1166. doi:10.1016/j.mri.2008.01.016
[2] C. H. Polman, et al., “Diagnostic Criteria for Multiple
Sclerosis: 2005 Revisions to the ‘McDonald Criteria’,”
Annals of Neurology, Vol. 58, No. 6, 2005, pp. 840-856.
doi:10.1002/ana.20703
[3] D. H. Miller, R. I. Grossman, S. C. Reingold and F. Mc-
Farlan, “The Role of Magnetic Resonance Techniques in
Understanding and Managing Multiple Sclerosis,” Brain,
Vol. 121, No. 1, 1998, pp. 3-24.
doi:10.1093/brain/121.1.3
[4] D. Yamamoto, et al., “Computer-Aided Detection of
Multiple Sclerosis Lesions in Brain Magnetic Resonance
Images: False Positive Reduction Scheme Consisted of
Rule-Based, Level Set Method, and Support Vector Ma-
chine,” Computerized Medical Imaging and Graphics,
Vol. 34, No. 5, 2010, pp. 404-413.
doi:10.1016/j.compmedimag.2010.02.001
[5] E. Geremia, et al., “Spatial Decision Forests for MS Le-
sion Segmentation in Multi-Channel MR Images,” Medi-
cal Image Computing and Computer-Assisted Interven-
tion, Vol. 6361, 2010, pp. 111-118.
[6] C. J. Wallace, T. P. Seland and T. C. Fong, “Multiple
Sclerosis: The Impact of MR Imaging,” American Jour-
nal of Roentgenology, Vol. 158, 1992, pp. 849-857.
[7] S. Wiebe, et al., “Serial Cranial and Spinal Cord Mag-
netic Resonance Imaging in Multiple Sclerosis,” Annals
of Neurology, Vol. 32, No. 5, 1992, pp. 643-650.
doi:10.1002/ana.410320507
[8] L. Truyen, “Magnetic Resonance Imaging in Multiple
Sclerosis: A Review,” Acta Neurologica Belgica, Vol. 94,
1994, pp. 98-102.
[9] F. Fazekas, et al., “The Contribution of Magnetic Reso-
nance Imaging to the Diagnosis of Multiple Sclerosis,”
Neurology, Vol. 53, 1999, pp. 448-456.
[10] P. Anbeek, K. L. Vincken and M. A. Viergever, “Auto-
mated MS-Lesion Segmentation by K-Nearest Neighbor
Classification,” MIDAS Journal, 2008.
http://hdl.handle.net/10380/1448
[11] F. Rousseau, F. Blanc, J. de Seze, L. Rumbach and J.
Armspach, “An a Contrario Approach for Outliers Seg-
mentation: Application to Multiple Sclerosis in MRI,” 5th
IEEE International Symposium, Paris, 2008, pp. 9-12.
[12] B. Johnston, M. S. Atkins, B. Mackiewich and M. And-
erson, “Segmentation of Multiple Sclerosis Lesions in
Copyright © 2011 SciRes. OJMI
B. A. ABDULLAH ET AL.
40
Intensity Corrected Multispectral MRI,” IEEE Transac-
tions on Medical Imaging, Vol. 15, No. 2, 1996, pp.
154-169. doi:10.1109/42.491417
[13] A. O. Boudraa, et al., “Automated Segmentation of Mul-
tiple Sclerosis Lesions in Multispectral MR Imaging Us-
ing Fuzzy Clustering,” Computers in Biology and Medi-
cine, Vol. 30, No. 1,2000, pp. 23-40.
[14] K. V. Leemput, F. Maes, D. Vandermeulen, A. Colches-
ter and P. Suetens, “Automated Segmentation of Multiple
Sclerosis Lesions by Model Outlier Detection,” IEEE Tr-
ansactions on Medical Imaging, Vol. 20, No. 8, 2001, pp.
677-688. doi:10.1109/42.938237
[15] A. P. Zijdenbos, R. Forghani and A. C. Evans, “Auto-
matic ‘Pipeline’ Analysis of 3-D MRI Data for Clinical
Trials: Application to Multiple Sclerosis,” IEEE Transac-
tions on Medical Imaging, Vol. 21, No. 10, 2002, pp.
1280-1291. doi:10.1109/TMI.2002.806283
[16] F. Kruggel, S. P. Joseph and H.-J. Gertz, “Texture-Based
Segmentation of Diffuse Lesions of the Brain’s White
Matter,” NeuroImage, Vol. 39, No. 3, 2008, pp. 987-996.
doi:10.1016/j.neuroimage.2007.09.058
[17] W. F. Liu, X. X. Zhou, G. L. Jiang and L. Z. Tong,
“Texture Analysis of MRI in Patients with Multiple
Sclerosis Based on the Gray-Level Difference Statistics,”
Education Technology and Computer Science, Vol. 3,
2009, pp. 771-774.
[18] A. Pozdnukhov and M. Kanevski, “Monitoring Network
Optimisation for Spatial Data Classification Using Sup-
port Vector Machines,” International Journal of Envi-
ronment and Pollution, Vol. 28, No. 3-4, 2006, pp. 465-
484.
[19] M. Kanevski, M. Maignan and A. Pozdnukhov, “Active
Learning of Environmental Data Using Support Vector
Machines,” Conference of the International Association
for Mathematical Geology, Toronto, 21-26 August 2005.
[20] R. R. Edelman, J. R. Hesselink, M. B. Zlatkin and J. V.
Crues, “Clinical Magnetic Resonance Imaging,” 3rd Edi-
tion, Elsevier, New York, 2006.
[21] R. Khayati, M. Vafadust, F. Towhidkhah and S. M. Na-
bavi, “Fully Automatic Segmentation of Multiple Sclero-
sis Lesions in Brain MR FLAIR Images Using Adaptive
Mixtures Method and Markov Random Field Model,”
Computers in Biology and Medicine, Vol. 38, 2008, pp.
379-390.
[22] W. I. McDonald, et al., “Recommendation Diagnostic
Criteria for Multiple Sclerosis: Guidelines from the In-
ternational Panel on the Diagnosis of Multiple Sclerosis,”
Annals of Neurology, Vol. 50, 2001, pp. 121-127.
[23] A. Younis, M. Ibrahim, M. Kabuka and N. John, “An
Artificial Immune-Activated Neural Network Applied to
Brain 3D MRI Segmentation,” Journal of Digital Imag-
ing, Vol. 21, Suppl. 1, 2008, pp. S69-S88.
[24] J. Rexilius, H. K. Hahn, H. Bourquain and H.-O. Peitgen,
“Ground Truth in MS Lesion Volumetry—A Phantom
Study,” Lecture Notes in Computer Science, Vol. 2879,
2003, pp. 546-553.
[25] C.-C. Chang and C.-J. Lin, “LIBSVM: A Library for
Support Vector Machines,” ACM Transactions on Intel-
ligent Systems and Technology, Vol. 2, No. 3, 2011, pp.
1-27.
[26] A. K. Jain, R. P. W. Duin a nd J. Mao, “Statistical Pattern
Recognition: Review,” IEEE Transactions Pattern
Analysis and Machine Intelligence, Vol. 22, No. 1, 2000,
pp. 4- 37. doi:10.1109/34.824819
[27] D. Garcia-Lorenzo, L. Lecoeur, D. Arnold and D. L.
Collins, “Multiple Sclerosis Lesion Segmentation Using
an Automatic Multimodal Graph Cuts,” Medical Image
Com- puting and Computer Assisted Intervention, London,
20-24 September 2009, pp. 584-591.
[28] D. Goldberg-Zimring, H. Azhari, S. Miron and A. Achi-
ron, “3-D Surface Reconstruction of Multiple Sclerosis
Lesions Using Spherical Harmonics,” Magnetic Reso-
nance Imaging, Vol. 46, 2001, pp. 756-766.
[29] Z. R. Chi, H. Yan and T. Pham, “Fuzzy Algorithms: With
Application to Image Processing and Pattern Recogni-
tion,” World Scientific Publishing Co. Pte. Ltd, Singa-
pore, 1996.
[30] J. Lecoeur, et al., “Multiple Sclerosis Lesions Segmenta-
tion using Spectral Gradient and Graph Cuts,” Proceed-
ings of MICCAI Workshop on Medical Image Analysis on
Multiple Sclerosis, New York, September 2008.
Copyright © 2011 SciRes. OJMI
B. A. ABDULLAH ET AL.
Copyright © 2011 SciRes. OJMI
41
Appendix A. Textural Features Extraction
Techniques
Textural features can be categorized according to the
matrix or vector used to calcu late the feature. In this sec-
tion, we are interested with histogram, gradient, run-
length matrix and co-occurrence based features. These
categories include features that are selected after being
tested to be identifying for the texture of regions that
suffer from the multiple sclerosis lesions. For all feature
calculations, the image is represented by a function f(x,y)
of two space variables x and y, x = 0,1,··· N – 1 and y =
0,1,···, M – 1. The function f(x,y) can take any value I =
0,1,···,G – 1 where G is total number of intensity levels in
the image.
Histogram Based Features
The intensity lev el h istogram is a function h(i) providing,
for each intensity level i, the number of pixels in the
whole image having this intens ity.
 


11
00
1,
,,; ,0,
NM
xy
ji
hifxyiijji





The histogram is a concise and simple summary of the
statistical information contained in the image. Dividing
the histogram h(i) by the total number of pixels in the
image provides the approximate probability density of
the occurrence of the intensity levels p(i), given by:
 
pihi NM
The following set of textural features is calculated
from the normalized histogram:
Mean:

1
0
G
i
ip i
Variance:

12
2
0
G
i
ip


i
Skewness:

13
330
1G
i
ip


i
Kurtosis:

14
440
13
G
i
ipi


Gradient Based Features
The gradient matrix element g(x,y) is defined for each
pixel in the image based on the neighborhood size. For a
3 × 3 pixels neighborhood, g is defined as follows:



22
1, 1,
,1 ,1
,
x
y
xy
f
xyfxy
fxy fxy
gxy
 
 

The following set of textural features is calculated
from the gradient matrix:
Mean of absolute gradien t
(GrMean) =

11
00
1,
NM
xy
xy
NM



Variance of absolute gradient
(GrVariance) =


11 2
00
1,
NM
xy
g
xy GrMean
NM



Skewness and kurtosis of the absolute gradient can be
calculated similar to those calculated for histogram.
Run Length Matrix Based Features
The run length matrix is defined for a specific direction.
Usually a matrix is calculated for the horizontal, vertical,
45° and 135° directions. The matrix element r(i,j) is de-
fined as the number of times there is a run of length j
having gray level i. Let G be the number of gray levels
and Nr be the number of runs. The following set of tex-
tural features is calculated from the run length matrix:
Short run emphasis inverse moments
(ShrtREmph) =
1
2
01
,
r
N
G
ij
rij C
j





Long run emphasis moments
(LngREmph) =

12
01 ,
r
N
G
ij
jr ijC





Gray level non-uniformity
(GLevNonUni) =

12
01,
r
N
G
ij
rij C









Run length non-uniformity
(RLNonUni) =

2
1
10 ,
r
NG
ji
rij C









where the normalization coefficient C is defined as fol-
lows: C =

1
01 ,
r
N
G
ij
rij


Co-Occurrence Matrix Based Features
The co-occurrence matrix is a form of second order his-
togram that is defined for certain angle θ and certain dis-
tance d. The matrix element hdθ (i, j) is the number of
times f (x1, y1) = i and f (x2, y2) = j where (x2, y2)=(x1, y1)
+ (dcosθ, dsinθ). Usually the co-occurrence matrix is
calculated for d = 1 and 2 with angles θ = 0°, 45°, 90° and
135°. When the matrix element hdθ (i, j) is divided by the
total number of neighboring pixels, the matrix becomes
the estimate of the joint probability codθ (i, j) of two pix-
els, a distance d apart along a given direction θ having
B. A. ABDULLAH ET AL.
42
co-occurring values i and j. Let µx, µy,
x
and
y
de-
note the mean and standard deviation of the row and
column sums of the matrix co, respectively. The follow-
ing set of textural features is calculated from the co-oc-
currence matrix:
Angular second moment
(AngScMom) =


11 2
00 ,
GG
ij
co ij



Contrast =

11 2
00 ,
GG
ij
ijcoij



Correlation =
11
00
,
GG
ij xy
ijco ijxy




Inverse Diff erence =


11
2
00
,
1
GG
ij
co ij
ij

 

Entropy =
 

1
2
01 ,log ,
r
N
G
ij
coij coij


Copyright © 2011 SciRes. OJMI