13th International Conference on Electronics Computer and Computation (ICECCO 2017 1

A Mobile-based Image Analysis System for

Cervical Cancer Detection
Saka Abiola Monsur,1∗ Steve A. Adeshina 2∗ Shivani Sud 3∗ and Winston O.Soboyejo 4∗
1
Department of Computer Science,Nile University of Nigeria, Abuja,
E-mail: saka.abiola@nileuniversity.edu.ng
2
Department of Computer Engineering, Nile University of Nigeria, Abuja,
E-mail: steve.adeshina@nileuniversity.edu.ng
3
University of North Carolina, U.S.A
4
Princeton University, Princeton, New Jersey, U.S.A

Abstract—Cervical cancer is the third major killer disease in developed cancer imaging research work. Additionally, the high cost
and developing countries. Whereas screening and other preventive of Colposcopy, which may have provided the specialist
measures reduce the mortality in developed countries, mortality rate still Physician an equivalent tool for diagnosis is a hinderance
remains very high in developing countries. This project focuses on the in developing countries.
analysis of digital image of the cervix, captured with a low-level camera
The closest work to our work is that of Mustapha et
and under a contrast agent (the visual inspection with acetic acid (VIA)).
Gaussian and mean filter techniques were used to remove the speckles.
al.where ’A low cost methodology in classifying cervical
A segmentation algorithm was used to isolate the Region of Interest cancer tissues for effective screening in developing coun-
(ROI) from the image. Additionally a Canny edge detection algorithm tries of the world’ [2]. This work was implemented with
was used to find edges. Furthermore, quantification and classification of ImageJ. Several processes in this approach were manually
the images were done. An android application was used to integrated implemented. Our work differs from the above in that,
all the above. This allows usage in rural settings. The results obtained whereas [2] implemented their system using several manual
were quite satisfactory (Specificity 79% and Sensitivity of 83%). processes and requires a dedicated server and computing
expert to operate, we have automated and integrated the
Index Terms—Cervical cancer, Region of Interest, Acetowhite region, entire procedure into a mobile application. The physician in
Human Papilomavirus, Visual Inspection.
a remote areas could use the system. Indeed an individual
could take a selfie of her cervix and get a diagnosis.
The innovation in this work is the integration and au-
tomation of several complex image processing and com-
1 I NTRODUCTION puter vision techniques into a mobile device. This will
Cervical cancer is one of the curable types of cancers in help a General Practitional (GP) or an individual to get a
women if detected early. Most cases of cervical cancer are prelimnary diagnosis in remote areas of the world.
caused as a result of infection with certain types of Human Figure 1 shows the high level process diagram showing
Papillomavirus (HPV) [4, 6]. Although women who have the various steps involved in the implementation of this
early exposure to sexual relationship and those with mul- work.
tiple sexual partners are at high risk of contracting HPV
and eventually, cervical cancer, it is however possible for a
woman to be infected with HPV even if she has had only
one sexual partner. In the developed nations, women above
the age of 30, who are at high risk of HPV infection are
given HPV vaccines, to reduce the chances of having the
disease [4]. In the developing countries, the lack of orga-
nized, population-wide screening programmes and medical
personnel to administer and translate various test involved
[6], makes the prevalence of Cervical Cancer endemic.
This research work seek to address the problem of the
lack of medical personnel to administer and translate var-
ious test by coming up with the automation of cervical
cancer detection using digital images of low resolutions,
which will be deployed on android mobile devices, an Fig. 1: Process diagram of the entire system
innovation we are not aware of in the history of cervical

978-1-5386-2501-9/17/$31.00 
c 2017 IEEE
 2

2 R ELATED W ORK
Other related work include that Yeshwanth Srinivasan et
al.[7] where they proposed a unified framework for a fully
automated system for the diagnosis of Cervical Intraepithe-
lial Neoplasia (CIN). In their approach, several algorithms
that are based on mathematical morphology, and clustering
based on Guassian Mixture Modeling (GMM) in a joint color
and geometric feature space were used to segment the micro
regions of the cervix (mosaicm, vasculature and punctuation
which made up the Acetowhite region [AW]). Not only
that these algorithms are of great importance in helping
expert to evaluate the inter-capillary distance, which is the a b
most important indicator of severe CIN, it also helps in
quantifying precisely, the extent of the abnormality. Fig. 2: (a) Conceptual diagram of the proposed system, (b)
Additionally Abhishek Das et al.[6] proposed the elim- Activity diagram
ination of specular reflection and the identification of the
ROI as the first steps in the automated detection of cer-
vical cancer using digital colposcopy. In their system, the
whole scheme was decomposed into 4 different modules.
Specular reflection removal (SR) from raw cervigram; seg-
mentation of cervix ROI; segmentation of cervix ROI into
Acetowhite(AW), columnar epithelium(CE) and squamous
epithelium(SE); classification of AW regions into AW, mo-
saic or punctuation tiles. K-means Clustering algorithm was
used in segmenting the cervix ROI from the raw image.
However, they were only able to present the first two-step
in their work.
Fig. 3: Class diagram of the proposed system

3 M ETHODS
3.1 Data Set
The images were sourced from an online image database;
Geneva Foundation for Medical Education and Research
(GFMER) [24]. A total of 24 classified images were used
(as a pilot test) in this research work. The classification was
made by VIA by experts in the field of oncology. Out of
these images, six images show no reaction with acetic acid,
15 indicate probably an early stage of cervical cancer (CIN I
and few CIN II) while the last three indicate a chronic stage
of the disease (CIN III and infiltrating cancer).
3.4 Specular removal
During the process of image acquisition, through a digital
camera, light falls on the uneven surface of the cervix,
and specular reflection appears in the resulting image. This
appearance is characterized with a bright spot, heavily
saturated with a white light. This bright spot may hinder
the image analysis process. A despeckle algorithm was im-
plemented in Android programming. The algorithm looks
for these heavy bright spots in the image and convolutes it
with the RGB values of the pixels surrounding it.

3.2 Conceptual Diagrams 3.5 Conversion to grayscale

An automated self-contained mobile system for cervical
cancer detection is proposed. The work is presented in
five different phase; Image Acquisition, Specular reflection
removal, conversion to grayscale, image segmentation, Edge
detection, image classification and result visualization. Fig-
ure 2a shows the conceptual diagram for the proposed
system, while Figure 2b shows the activity diagram of the
system
The Class diagram of the designed system is shown in 3
3.3 Image Acquisition
The system has been designed in a way that allows two
different sources of image sample; Images can be uploaded
from the device storage into the system or a fresh image
of the cervix can be taken directly with a device with high
resolution and the use of speculum to magnify the image
after the application of Acetic Acid to the surface area of the
cervix.
For better manipulation and efficient use of computing
resources, it is better to operate with a greyscale image,
rather than a coloured image. The image is converted into
greyscale using various Java classes like Pain, ColorMatrix,
Canvas, Bitmap etc.
GrayScale = (Red ∗ 0.3 + Green ∗ 0.59 + Blue ∗ 0.11) (1)
Equation 1 was used for grayscale conversion.
3.6 Image Segmentation
Most medical images are accompanied with irrelevant infor-
mation which can affect the analysis process. It is therefore
pertinent to separate these ROI images. Many techniques
and algorithms have been discussed in the literature. These
techniques include Region-based, Thresholding, Edge-based
techniques, feature and model-based segmentation [7,8],
 3

while several algorithms have also been implemented: K-
Means Clustering algorithm, Watershed algorithm among
others. Thresholding technique was adopted for the purpose
of this research. Given a single threshold, t , the pixel located
at lattice position (i,j), with greyscale value fij is allocated
to category 1 if
fij <= t
Otherwise, the pixel is allocated to category 2. Category
1 represents the ROI while category 2 is the irrelevant part
of the image or the image background. T was chosen by
manually trying out several values. Threshold value of 95
happens to perform optimally at identifying the object of
interest. Each pixel in the other category is converted to
black.
Given Equation 2
T = T [x, y, p(x, y), f (x, y)] (2)
Where T is the threshold, x&y are the image coordinates,
p(x, y) and f (x, y) are points the gray level image pixels.
The threshold image can be defined as g(x, y) as in
Equation 3 ;
 24 images that formed our data set were first despecked as
1 if f (x, y) > 1
g(x, y) = −−−−−− (3)
0 if f (x, y) ≤ 0
3) Perform a non-maximum suppression; and
4) Trace edges with hysteresis.
This was implemented in Android programming.
3.8 Image Classification
There is a great relationship between the frequency of edges
in image and the possibility of cervical cancer occurrence
[2]. The mean grey value of the resulting edges in the
image is calculated. From these values, mean and standard
deviations are also calculated. A threshold value is chosen
whereby any image that resulted in a mean below or equal
to the threshold value is adjudged to be free of cancer, while
the one just a little above the threshold value indicates a
great possibility of the presence of cervical cancer at its
early stage; CIN1 . Lastly, any image with a mean value
of far above the threshold value has high tendencies of the
presence of chronic cancer, CIN2 or CIN3.
4 EXPERIMENTS AND RESULTS
shown in Figure 4, and then converted to grayscale for easy
processing, as soon in the Figure .

3.7 Edge Detection

After successful image segmentation, we look out for the
presence of edges in the ROI. This is known as edge
detection. Edge detection is a technique defining a set of
mathematical methods which aim at identifying points of
discontinuity in a digital image. In digital image processing,
several features can be chosen in classifying the image. This
includes intensity of the image and the edges in images. This a b
approach is one of the best for cervical cancer imaging. Sev-
eral edge detection algorithms have been implemented, for Fig. 4: (a) Original image, (b) After specular removal
example canny edge, Sobel, Prewitt, and Kirsch [9] amogst
others. These can be largely categorized under Laplasian
techniques and Gaussian techniques. Canny Edge detection
algorithm was considered the most approprite for applica-
tions in this category [9]. After series of experimentation
edges detected by Canny were found to correllate with
blood marking abrasions on the cervix.
Amidst all the aforementioned edge detection methods,
canny edge has been rated as the best [9]. Hence, it was
adopted for this work. a b
The procedure involved in the canny edge detection is
as follows; Fig. 5: (a) Image without Specular removal, (b) Its greyscale
equivalent
1) Smoothen the image to reduce noise, using Gaus-
sian filtering; The equation for a Gaussian filter of
kernel size (2k + 1)(2k + 1) is given by equation 4
 
2 2 4.1 Image Segmentation
1 − (i−(k+1)) 2σ
+(j−(k+1))
Hi,j = 2
e 2
(4) A despeckled image that has been converted to grayscale
2πσ
was passed to the image segmentation module. The ROI was
2) Compute the intensity gradient as in Equation 5 ;
 mapped out of the whole image leaving out the background.
G = G2x + G2x , θ = arctan 2(Gx , Gy ) (5) Threshold gray value of 95 was used in identifying the
ROI. This was determined after experimenting with several
Where Gx and Gy represent the first derivative of values. Figure 6 shows the despeckled grayscale image and
the horizontal and vertical direction respectively ROI.
 4

a b

Fig. 6: (a) Original image, (b) ROI segmentation

4.2 Edge Detection

We used canny edge for the edge detection. The parameter
used for the canny edge algorithm implementation includes;
low threshold of 2.5, high threshold of 7.5, guassian kernel
radius of 3, guassian kernel radius of 18. Figure 7 shows the
resultant edges in one of the images we experimented with. a b

Fig. 8: (a) Input Screen (b) Stored Images

a b

Fig. 7: (a) ROI image (b) Edges in the image

4.3 Image Classification

After several trials of experiment, we arrived at certain
threshold for the mean gray values of the edges; values
below are classified as free from cancer, values between 7
and 9 are classified as CNI (cancer at its early stage) and a b
finally, a value of above 9 is classified as CNII (infiltrating
cancer). Fig. 9: (a) Detection process (b) Analysis
The experimental result was confirmed with the ground
truth value of the data set used. It was quite satisfactory as
soon on the performance and measurement section. 4.6 Discussions and Conclusions
The result obtained using the 24 images in the dataset is
4.4 Mobile Applications working process summarized using bar chart as shown in Figure 11.This
result is comparable to what obtains in the literature, though
Figure 8 shows the input screen of the mobile application we are not aware any implementation in a mobile device.
while 9 shows the detection and analysis processes. This novel approach of packing all the activities of assess-
ment into a mobile device will go a long way in reducing
the number of deaths arising from cases of cervical cancer.
4.5 Performance and Measurement
Since it will aid early detection of the condition, which is the
The table below shows the performance of our system in main key to surviving the dreadful condition. In addition,
term of the specificity and the sensitivity. this simple tool can be used in rural settings where there are
no medical experts.
Figure 10 shows that, there are three conditions of cer-
No of TP 8 No of TN 11
No of FN 2 No of FP 3 vical cancer cases; Firstly, if the analysis resulted in a low
Sensitivity 83 % Specificity 79 % mean grey value, it means there are fewer edges in the
image; therefore, there is no case of cervical cancer. Secondly,
TABLE 1: Systems performance if the mean grey value is just a bit above the threshold value,
this could indicate a trace of cervical cancer in its early stage.

a b
Fig. 10: (a) Result page - Positive (b) Result page - Negative
Fig. 11: Result summary
However, there is a possibility of this being a False Positive,
resulting from some conditions in the sample image such as;
camera reflection, glare from acid pool, darkness from blood
on the wall of cervix etc. In this case, a confirmatory test like
biopsy can be suggested. Finally, a high mean grey value
relative to the threshold value, as a result of multiple edges
from; the proliferated, bulging or mushroom-like growth
on the surface of the sample image indicates a high grade
cervical cancer (CIN III or Infiltrating Cancer) and thus,
requires a very urgent medical attention.
This work has resulted in a portable android mobile
application for the analysis of cervical cancer images. This
system is limited to android devices and some hybrid ver-
sions of blackberry devices, thus leaving out the iOS devices
user.
The number of images is a limitation of this work, how-
ever using a bottom-up approach (edge detection) helped
in getting reasonable results. It is hoped that in future a
conversion of the codes will be done through the new ’ionic
framework’. This is so that the application is not limited
to Android devices alone. It will be available for iOS and
other devices. Furthermore, it will desirable to impement
this application using model based approach such as Convo-
lutionary Neural Networks. This will require several images
and the processes to acquire more images is in progress.
5
ACKNOWLEDGMENTS
We wished to acknowledge the Pan African Material Science
Institute.
R EFERENCES
[1] American Cancer Society Cancer Facts and Figures
2015. Atlanta: American Cancer Society; 2015 Available at
/www.cancer.org/acs/groups/content/@editorial/.../acspc-044552.pdf
[2] Suleiman Mustafa, Steve Adeshina, Mohammed Dauda and Wole
Soboyejo Classification of cervical cancer tissues using a novel low
cost methodology for effective screening in rural settings. Proc.
IEEE International Conference on Electronics Computers and Computa-
tion 2014
[3] World Health Organization Organization Global Health Observa-
tory Data Repository, Mortality and Global Health Estimates 2012
[4] Abhishek Das, Avijit Kar and Debasis Bhattacharyya; Detection
of abnormal regions of precancerous lesions in Digitised Uterine
Cervix images In IEEE, ISBN 978-1-4799-3174-3/14, 2014
[5] Usman Arshad, Cecilia Mascolo and Marcus Mellor Exploting
mobile computing in health-care IWSAWC, 2014
[6] Abhishek Das, Avijit Kar and Debasis Bhattacharyya Elimination
of Specular reflection and identification of ROI: The first step in
automated detection of cervical cancer using digital colposcopy
IEEE, ISBN 978-1-61284-896-9/11, 2014.
[7] Yeshwanth Srinivasan,Enrique Corona, Brian Nutter,Sunanda Mi-
tra and Sonal Bhattacharya, A Unified Model-Based Image Analysis
Framework for Automated Detection of Precancerous Lesions in
Digitized Uterine Cervix Images IEEE JOURNAL OF SELECTED
TOPICS IN SIGNAL PROCESSING, VOL. 3, NO. 1, FEBRUARY 2009
[8] A. F. Zuluaga, M. Follen, I. Boiko, A. Malpica, and R. Richards-
Kortum, Optical coherence tomography: A pilot study of a new
imaging technique for noninvasive examination of cervical tissue.
American journal of obstetrics and gynecology, vol. 193, pp. 83-88,
2005.
[9] Raman Maini and Dr. Himanshu Aggarwal Study and comparism
of various image edge detection techniques, International Journals
of Image Processing, Vol 1,
[10] Isaac N Bankman Handbook of medical imaging processing and
analysis, Academic press, ISBN 0-12-077790-8, 2000.
[11] Gonzalez, R. C., and Woods, R. E Digital Image Processing 3rd
Edition Addison-Wesley, Reading, MA, 1992.
[12] M.Clark, A. Bovik, and W. Geisler, Texture segmentation using a
class of narrow band filters, IEEE, 1987.
[13] K. Hammouda and Jernigan Texture Segmentation using Gabor
Filter.
[14] Programming in Java Advanced Imaging, Release 1.0.1, Novem-
ber 1999.
[15] Android development training; available at Android development
training; available at
[16] John W. Sellors and R Sankaranarayanan, Colposcopy and Treat-
ment of Cervical Intraepithelial Neoplasia: Beginner’s guide, IARC
Press, 2013, ISBN 92 832 0412 3.
[17] Michele Follen and Charles F. Levenback, Imaging in Cervical
Cancer, American Cancer Society, 2003.
[18] Patel, Shilpa et al. Imaging of Endometrial and Cervical Cancer.
In Insights into Imaging 1.5-6 (2010): 309328. PMC. Web. 8 May 2016.