Content uploaded by Sandy Rihana
Author content
All content in this area was uploaded by Sandy Rihana on Feb 26, 2014
Content may be subject to copyright.
Automated Algorithm for Ovarian Cysts Detection
in Ultrasonogram
Sandy RIHANA
(1)
Hares Moussallem
(2)
Chiraz Skaf
(2)
Charles Yaacoub
(2)
(1)
Biomedical Engineering Department
(2)
Telecommunications Engineering Department
Holy Spirit University of Kaslik (USEK)
Jounieh, Lebanon
sandyrihana@usek.edu.lb
Abstract— Polycystic Ovary Syndrome (PCOS) is a female
endocrine disorder which severely affects women’s health
and its diagnostic requires medical treatment or even
surgery. Manual analysis of PCOS diagnosis often
produces errors. Recently, many automated algorithms
have been studied for polycysts detection in Ultrasound
images. This paper presents cysts detection and
classification in the ovary ultrasound images with an
accuracy that reaches 90%.
Keywords—ultrasound medical imaging, cysts,
thresholding, multiscale morphological method , svm
I.
I
NTRODUCTION
Diagnostic ultrasound (US) is nowadays the most common
noninvasive medical-imaging modality. In fact, the first step in
the roadmap for the diagnostic of ovarian cystic masses is
based on ovarian ultrasound. Ovarian ultrasounds are
maneuvered by gynecologist in order to detect and heal cysts
that may occur.
These cysts are developed due to incomplete developed
follicles in the ovaries. They can generally be detected on the
ultrasound images by some dark regions, darker than other
regions in the same image, thus tracing a sort of edge of an
elliptic geometric shape.
Manual analysis by clinicians is generally used in the
diagnostic. Generally, periodic measurements of the size,
texture and shape of follicles over several days are the primary
means of evaluation. Nevertheless, nowadays, automated
software able to help the clinicians to identify the cysts and to
reduce the burden of the clinical diagnosis in order to
differentiate among malignant and benign cysts could be
appropriate. The objective of this paper is to develop an
extended processing scheme for automatic detection of
follicles in ultrasound images of ovaries.
Different methods have been developed in the literature on the
identification of the ovarian cysts. Potocnik and Zazula worked
on a method based on optimal thresholding [1] and then
upgraded by using active contour technique [2] for the
segmentation of follicles. Cigale and Zazula implemented the
neural network approach for the segmentation [3]. Others use
the multiscale morphological method for the denoising,
contrast enhancement [4], and horizontal and vertical
thresholding for cysts segmentation [5] .
The method presented in this paper consists of applying a
multi-scale morphological process for noise reduction and for
the contrast enhancement, followed by segmentation and
detection. It aims at extracting features that every clinician
bases his diagnostic on. These parameters are major axis,
minor axis, area and perimeter. In addition, it allows us to
differentiate between simple, polycysts and endometrioma
cysts, by calculating the mean and the standard deviation of a
sub-image extracted from each detected cyst.
II.
M
ATERIALS AND METHODS
US image acquisition
Digital recordings of ovaries ultrasound scans have been
provided by the obstetrics gynecology polyclinic MD. Barakat.
These images were performed on a group of women with no
cysts, simple cysts and polycysts syndrome. Some of the
simple cysts syndromes present endometrioma cyst [1]. These
images were assessed by the clinician through periodic
examinations and some of them through surgery. At total, 25
ultrasound images are used for the development of this method.
US image processing
Figure 1 shows the image processing flowchart developed
in this paper. A preprocessing part consists of contrast
enhancement of the grayscale image and in image binarization.
A post processing part consists of detecting and labelling
connected components leading finally to geometrical features
extraction and classification of the cysts.
A. Image preprocessing
After converting the images into grayscale, contrast
enhancement is performed based on morphological operation,
such as top hat and bottom hat. The former top hat, returns the
image minus the morphological opening of the image (erosion
followed by dilation) while the bottom hat transformation
returns the image minus the morphological closing of the
image (dilation followed by erosion). Equation 1 summarizes
the process:
2013 2nd International Conference on Advances in Biomedical Engineering
978-1-4799-0251-4/13/$31.00 ©2013 IEEE 219
1 1
( , ) ( , ) 0.5 ( , ) 0.5 ( , ),
m m
T B
iS iS
i i
g r c g r c F r c F r c
= =
= + −
∑ ∑
%
(1)
where ğ(r , c) the output pixel at coordinates (r , c) , S is a disk-
Figure 1- General bloc diagram
shaped structuring element of radius R=3 empirically used for
morphological opening and closing, F
T
iS
and F
B
iS
the top hat
and bottom hat transforms at scale i, containg respectively
bright and dark features smaller than S, and g the input image.
In a medical ultrasound image many undesired structures blur
the desired outcome of the image like blood vessels, nerve
fibers, lymphatic glands and added noise due to the ultrasound
waves propagation. So the detection of cysts becomes a
challenging task in such a noisy image. Therefore, traditional
edge based techniques (such as sobel, prewitt) give false results
when applied on these images due to added noise. The follicle
appears as a homogenous region in the ultrasound image. The
gray level values for the pixel within the follicles will be more
or less the same as the background. The thresholding method
proposed is based on horizontal and vertical scanline
thresholding respectively then merging both results in order to
obtain the binary image. The horizontal scanline thresholding
could be described as follows. The image g of size M x N is
considered. The mean m
r
and standard deviation σ
r
of the r
th
row sub image are given by equations (2) and (3), respectively.
( )
1
1
,
N
r
c
m g r c
N
=
=
∑
(2)
( )
( )
2
1
1
,
N
r r
c
g r c m
N
σ
=
= −
∑
(3)
A threshold T
2
= K
2
σ
r
is applied to the image for the
binarization, where K
2
is a scale multiplication factor. Same
procedure is applied for each column as expressed in equations
(4) and (5), where K
4
is the scaling factor and T
4
= K
4
σ
r
.
( )
1
1
,
M
c
r
m g r c
M
=
=
∑
(4)
( )
( )
2
1
1
,
M
c c
r
g r c m
M
σ
=
= −
∑
(5)
The results of horizontal and vertical scan-line thresholding
are then merged by applying the logical “OR” operation to
yield at the end a binary image mask.
B. Image postprocessing
Applying morphological opening on the binary image
enhances the quality of the image obtained by removing the
undesired small components. The morphological opening
operation is an erosion followed by a dilation, using the same
structuring element for both operations. The structuring
element used is disk-shaped element with radius R=3.
The region of interest (ROI) window is covered by waves
emitted by the US probe. The use of a mask having the same
shape as seen in Figure 3 limits the ROI and helps reducing
false detection. The mask is generated with adobe Photoshop
CS version by assigning to the ROI area a white color (i.e.
logic ‘1’) and the affecting the remaining area with black (i.e.
logic ‘0’). The binary image mask obtained after
morphological opening and the mask specifying the ROI are
then multiplied in order to better detect connected components
and filter out undesired image areas.
Figure 2- ROI Mask of same size of the original image
representing the angle of the probe
Generally, two adjacent pixels, assigned with similar values
belong to the same component. 8-connectity is used to
elaborate all connected components in the image. Connected
components within a certain range of size assumed to be
probable follicles or cysts are considered in the next steps.
C. Feature extraction
The medical diagnosis for identifying the cyst is based on
indicators such as the number of follicles exceeding a certain
size and their relative position in the ovary. Measurement is
mandatory and periodical, even daily analysis could be
performed over 8-10 days, depending on the situation. To
efficiently characterize follicles, some parameters should be
known to be compared with standard parameters. Geometrical
and texture features of the ovarian cysts in ultrasound images
such as area, major axis length, minor axis length, major axis
length to minor axis length ration, compactness extent, centroid
220
and so on helps characterizing these follicles. Following the
clinical flowchart of the clinician, the Area, the Major axis
length, the Minor axis length, and the centroid are extracted as
geometrical features. Area is the number of pixels included
inside the potential follicle. A circular form at a diameter
normally between 2 mm and 30 mm gives an area of 4mm
2
to
700mm
2
. This factor is crucial in differentiating between a
follicle and a cyst. Using the resolution (DPI) of the ultrasound
machine, the area in the metric system is calculated. Knowing
the major axis and the minor axis, the area is computed (6).
Area = π • (Major Axis • Minor Axis) (6)
Major and minor axis: all follicles and cysts have an oval shape
close to an elliptic form, and are therefore modeled by an
ellipse. The major axis and the minor axis are the ones
corresponding to the ellipse having the same 2
nd
central
moment as the segmented area (follicle or cyst). The centroid is
the center of mass of the region of interest characterized by its
x and y coordinates. Objects of interest in an ultrasound are
better visualized when they are in the middle of the field of
view (angle of the probe).
Geometrical features extraction orients the diagnosis
toward the absence or the presence of cysts based on its
location, shape, area… Nevertheless, it does not give a clue
about the type of the cyst, whether it is an endometrioma cyst
or a normal one. This differentiation will be done using texture-
based feature extraction. Figure 3 illustrates the difference in
texture between a normal cyst and an endometrioma cyst.
Figure 3- Normal (left) and Endometrioma (right) cysts
To be able to differentiate between these two types of cysts,
a sub-image of 17x17 is extracted from each identified region,
centered at the centroid, thus getting sure that the processing is
inside the ROI, then the mean and the standard deviation are
computed for each type of cyst. These two parameters are
different for dissimilar types and provide a reliable parameter
to classify the two types of cysts based on the texture.
D. Classification and validation using ROC
Linear Separate vector machine is used as a classification
method. The classifier takes the mean and the standard
deviation as input vector and gives an output differentiating
between normal and endometria cysts. . SVM is easily
implemented comparing to “Neural Network” or other
classifiers: Phase1- training: Inputs to the SVM, 2 vectors; an
input training feature vector and a class vector as the output
corresponding to the input. Phase 2- testing: a test vector as
input to validate the SVM classifier. This classifier is usually
reliable in differentiating between only 2 classes having no
interference. To validate the algorithm cited above, the
“Receiver Operating Characteristic” (ROC) analysis is a
common means of comparing precision, accuracy, and
efficiency. This method shows a good evaluation with simple
and clear criteria, used by the “American Statistical
Association” in the medical field. A population has been
chosen divided into 4 parts; True negative (TN), false negative
(FN), false positive (FP) and true positive (TP). There are two
potentials of errors: FP and FN; either the individual is non-
diseased with positive test or diseased with negative test. TP is
for diseased individual with positive test and TN is for non-
diseased individual with negative test. 80 images were divided
into 4 groups equally, between simple, poly, and endometrioma
cysts in addition to 20 healthy non-containing cysts. The
accuracy is defined as the capability of giving the right choice
without distinguishing between positive and negative.
100%
TP TN
Accuracy
TP FP TN FN
+
= ×
+ + +
(7)
The sensitivity is defined as the proportion of patients
having a disease and detected by the algorithm.
100%
TP
Sensitivity
TP FN
= ×
+
(8)
The specificity is defined as the portion of people diagnosted
free from any disease and in the same time they are not ill.
100%
TN
Specificity
TN FP
= ×
+
(9)
Figure 5 shows the importance of top hat and bottom hat
filter transformations and how the contrast of the dark regions
is enhanced. Comparing to Figure 4, the edge of the cysts is
getting more evident.
We consider next another ultrasonogram and show the results
of the binarization process. Figure 6 shows the result of
vertical scanline thresholding, and Figure 7 the horizontal
scanline. Both images are merged in Figure 8 using the logical
“OR” operator. The next step is to find the connected
components based on 8-connectivity; Figure 10 shows the
connected components detected and labeled. The only cyst is
labeled 1 and the other connected components were labeled as
2, 3, 4, 5 and 6 on the black background. These small groups
of white pixels should be removed in order to decrease the
probability of false cysts detections. After applying the
constraint on the cyst’s size and location based on the area,
centroid and location parameters, Figure 1 and 12 illustrate the
resulting segmented ultrasonogram of a simple cyst, and
polycyst (bi-cyst) US respectively.
For texture based classification, the mean and the standard
deviation of the segmented cysts characterize the image.
Twenty simple cysts images and twenty endometrioma cysts
were taken for training phase. Another 5 images of each type
were used for the testing phase. SVM classifier is used and the
Figure represents the classified cysts.
221
Figure 4-Grayscale ultrasonogram Figure 5 - After contrast enhancement
Figure 6-Horizontal scanline and Vertical scanline thresholding
Figure 7- Vertical and Horizontal merged the logical OR operator
As we see the simple cyst class is found above the kernel
function and the endometrioma are below the kernel function.
To validate this algorithm, the accuracy, sensitivity and
specificity have been computed :
accuracy= 90% , sensitivity= 88.33%, specificity= 95%
Figure 10- (a) Result of morphological opening, (b) ROI mask, (c) negative of
the result obtained by multiplying image in (a) with the one in (b).
IV.
C
ONCLUSION
This paper presents the elaboration of a new algorithm capable
of detecting the cycts in ovarian ultrasound images and of
differentiating between the two types of cysts. This detection
and classification has been made based on the geometrical
features of the cysts their texture. The accuracy found of 90%
is a promising result. For future work, the algorithm would be
improved to generalize the work on all ovarian cyst types,
including Dermoid cysts, while keeping in mind the
improvement of the classification accuracy. Increasing the
database size and the number of images would therefore be
necessary for a better evaluation of the solution.
Figure 11- Connected components found and labeled
Figure 12- Simple cyst segmented
Figure 13- Bi-cyst segmented and labeled
Figure 14- Graph representing simple and endometrioma classes separated by
the linear kernel function
A
CKNOWLEDGMENT
The authors would like to address their special thanks to MD.
Habib Barakat, who provided the ultrasonograms and the
clinical assessment of each ultrasound image.
R
EFERENCES
[1] A.H.Balen, J.S.E.Laven, S-L.Tan, D.Dewailly, "Ultrasound assessment of
the polycystic ovary”, international consensus definitions, Hum Reprod Update
2003; 9:505-514.
[2] Cigale B and Zazula D, "Segmentation of ovarian ultrasound images using
cellular neural networks," in Proc. 7th Inter .work. systems, signals and image
processing, 2000.
[3] Anthony Krivanek and Milan Sonka, "Ovarian Ultrasound Image Analysis:
Follicle Segmentation," IEEE Transaction on medical imaging, vol. 17, pp.
935-944, 1998.
[4] P.S.Hiremath and Prema Akkasaliger, "Despeckling medical ultrasound
images using the contourlet transform," in 4th AMS Indian International
Conference on Artificial Intelligence, 2009.
[5] P.S.Hiremath and J.R.Tegnoor, "Automated detection of follicle in
ultrasound images of ovaries using edge based method," Recent trends in image
processing and pattern recognition, pp. 120-125, 2010.
222
