International Journal of Trend in Scientific Research and Development (IJTSRD)

International Open Access Journal | www.ijtsrd.com

ISSN No: 2456 - 6470 | Volume - 2 | Issue – 6 | Sep – Oct 2018

Skin Cancer Detection using Digital Image Processing and

Implementation using ANN and ABCD Features
Khaing Thazin Oo1, Dr. Moe Mon Myint2, Dr. Khin Thuzar Win3
1
Assistant Lecturer, 2,3Professor
1,2
Department of Electronics Engineering, 3Department of Mechronics Engineering,
Engineering
Pyay Technological University, Myanmar

ABSTRACT Keyword: Skin n cancer, Segmentation, Feature

Melanoma is a serious type of skin cancer. It starts in Extraction, Classification, Melanoma
skin cells called melanocytes. There are 3 main types
of skin cancer, Melanoma, Basal and Squamous cell I. INTRODUCTION
carcinoma. Melanoma is more likely to spread to Skin cancers can be classified into melanoma and
other parts of the body. Early detection of malignant non-melanoma.
melanoma. Melanoma is a malignancy of the
melanoma in dermoscopy copy images is very important cells which gives the skin its colour
colo (melanocytes)
and critical, since its detection in the early stage can and it can invade nearby tissues. Moreover, it spreads
be helpful to cure it. Computer Aided Diagnosis through the whole human body and it might cause to
systems can be very helpful
elpful to facilitate the early patient death and non-melanoma
melanoma which is rarely
detection of cancers for dermatologists. Image spread to other parts of the human body. Malignant
processing is a commonly used method for skin melanoma is the most aggressive type of human skin
cancer detection from the appearance of affected area cancers and its incidence has been rapidly increasing
on the skin. In this work, a computerised method has [1] [2] [3] [4]. Nevertheless, it is also the most
been developed to make usee of Neural Networks in treatable type of skin cancer if detected or diagnosed
the field of medical image processing. The ultimate at an early stage [5]. The diagnosis of melanoma in
aim of this paper is to implement cost cost-effective early stage is a challenging and fundamental task for
emergency support systems; to process the medical dermatologists since some other skin lesions may
images. It is more advantageous to patients. The have similar physical characteristics. Dermoscopy
Dermos is
dermoscopy image of suspect area of skin cancer is considered as the widely common technique used to
taken and it goes under various pre pre-processing perform an in-vivo
vivo observation of pigmented skin
technique for noise removal and image enhancement. lesions [6]. In early detection
n of malignant melanoma,
Then the image is undergone to segmentation using dermoscopic images have great potential, but their
Thresholding
sholding method. Some features of image have to interpretation is time consuming and subjective, even
be extracted using ABCD rules. In this work, for trained dermatologists. Therefore, the need to
Asymmetry index and Geometric features are build a system which can assist dermatologists to get
extracted from the segmented image. These features right decision for their
eir diagnosis has become very
are given as the input to classifier. Artificial Neural important.
Network (ANN) with feed forward architecture is
used for classification purpose. It classifies the given Image processing is one of the widely used methods
image into cancerous or non-cancerous.
cancerous. The proposed for skin cancer detection. Dermoscopy could be a
algorithm has been tested on the ISIC (International non-invasive
invasive examination technique supported the
Skin Imaging Collaboration) 2017 training and test cause of incident light beam and oil immersion
datasets. The ground truth data of each image is technique to form potential
otential the visual investigation of
available as well, so performance of this wor work can surface structures of the skin. The detection of
evaluate quantitatively. melanoma using dermoscopy is higher than individual
observation based detection [3],
[ but its diagnostic

@ IJTSRD | Available Online @ www.ijtsrd.com | Volume – 2 | Issue – 6 | Sep-Oct

Oct 2018 Page: 962
 International Journal of Trend in Scientific Research and Development (IJTSRD) ISSN: 2456-6470
2456
accuracy depends on the factor of training the computerised melanoma detection using Artificial
dermatologist. Neural Network classification has been adapted which
is efficient than the conventional one and Melanoma
The diagnosis
iagnosis of melanoma from melanocytic nevi is detection using Artificial Neural Network is a more
not clear and easy to identify, especially in the early effective method compared to other.
stage. Thus, automatic diagnosis tool is more effective
and essential part of physicians. Even when the II. Methodology
dermoscopy for diagnosis is done with the expert The following steps are implemented for classification
dermatologists,
ermatologists, the accuracy of melanoma diagnosis of skin cancer.
is not more than 75-84% [4]. 4]. The computer aided
Image Acquisition
diagnostics is more useful to increase the diagnosis
Pre-processing
accuracy as well as the speed [5].
Segmentation

Feature Extraction
The computer is not more inventive than human but
Classification
probably it may be able
le to extract some information,
like colour variation, asymmetry, texture features, The first step is the capture image the skin lesion
more accurately that may not be readily observed by image is acquired from the ISIC ISI 2018 database
naked human eyes [5]. 5]. There have been many through MATLAB. The second step is image pre- pre
proposed systems and algorithms such as the sseven- processing, the pre-processing
processing technique can be
point checklist, ABCD rule, and the Menzies method applied to eliminate the irrelevant data contains in the
[2, 3]] to improve the diagnostics of the melanoma image. The skin lesion regions such as label, marks,
skin cancer. oils, hairs are removed from the original image using
median filter techniques. The third step is image
The key steps in a computer-aided aided diagnosis of segmentation: the goal of image segmentation is to
melanoma skin cancer are image acquisition of a skin make simpler change the representation of an image
lesion, segmentation of the skin lesion from skin into something that is more meaningful
meaning and easier to
region, extraction of geometric features of the lesion analyze. The next step is image enhancement; to
blob and feature classification. Segmentation or improve the quality of the image so that the
border detection is the course of action of separating consequential image is better than the original image.
the skin lesion of melanoma from the circumferential In this step, the required region is segmented out and
skin to form the area of interest. Feature extraction is detected the edge of the skin lesion.
lesio The fourth step is
done to extract
xtract the geometric features which are feature extraction and the final step is classification of
accountable for increasing the accuracy; the skin lesion image.
corresponding to those visually ally detected by
dermatologists, that meticulously characterizes a
melanoma lesion.

The feature extraction methodology of many

computerised melanoma detection systems has been
largely depending on the conventional clinical Fig. 1System
System Block Diagram image
diagnostic algorithm of ABCD-rule rule of dermoscopy
due to its effectiveness and simplicity of A. Pre-processing
implementation [7].]. The effectiveness of methodology In this work, few previous processing techniques are
stems from the fact that it incorporates tes the classic needed. Prior to segmentation, all images are pre-
features of a melanoma lesion such as asymmetry, processed in order to minimize undesirable features
border irregularity, colour and diameter (or that could affect the performance of the algorithm
differential structures), where surveyable measures such as reflections, presence of the hair and colour
colo
can be computed. differences between images. The image is also
normalized to a unique size and shape. This
Dermoscopy is a diagnostic technique that is used normalization allows for the comparison of the region
worldwide in the recognitionn and interpretation of features such as positions and sizes between different
copious skin lesions [4]. Other than dermoscopy, a images.

@ IJTSRD | Available Online @ www.ijtsrd.com | Volume – 2 | Issue – 6 | Sep-Oct

Oct 2018 Page: 963
 International Journal of Trend in Scientific Research and Development (IJTSRD) ISSN: 2456-6470
2456
B. Skin Lesion Segmentation 1. Area (A):
A): Number of pixels of the lesion.
Image segmentation is an essential process for most 2. Perimeter (P): Number of pixels along the
image analysis subsequent tasks. Segmentation detected boundary
divides an image into its constituent regions or 3. Greatest Diameter (GD): ): The length of the line
objects. The goal of segmentation is to make simpler which connects the two farthest
or change the representation of an image into 4. Shortest Diameter (SD): The length of the line
something that is more meaningful and easier to connecting the two closest boundary points and
analyse. passes across the lesion centroid.
5. Circularity Index (CRC): It explains the shape
Image segmentation is the course of action of uniformity.
segregating an image into multiple parts, which is 6. Irregularity Index A (IrA):
used to identify objects or other relevant information 7. Irregularity Index B (IrB):
in digital images. Background subtraction, also known 8. Irregularity Index C (IrC):
as blob detection, is an emerging technique in the 9. Irregularity Index D (IrD):
fields of image processing wherein an image’s Diameter: Diameter in pixels
foreground is extracted for further processing.
Typically, an image’s regions of interest are objects in E. Classification
its foreground. Segmentation methods can be The main issue of the classification task is to avoiding
classified as thresholding, region based, Edge based over fitting caused by the small
sma number of images of
and clustering. In this work, thresholding is used skin lesion in most dermatology datasets. In order to
because of the computationally
omputationally inexpensive and fast solve this problem, the objective of the proposed
and simple to implement. model is to firstly extract features from images and
secondly load those extracted representations on an
C. Post-processing ANN network to classify.
Once binary image of skin lesion has been obtained,
the image then needs to be post processed in order to III. Performance Evaluation Parameter
resolve any problems there may be as follows: The most common performance measures consider

To prevent edges which are dark in many of the the model’s ability to discern one class versus all
images, from remaining as part of the lesion mask; others. The class of interest is known as the positive

To reduce any effects of disturbing artifacts as far class, while all others are known as negative. The
as possible; relationship between en positive class and negative class

To set soft edges, to ensure there are not too many predictions can be depicted as a 2  2 confusion
recesses or projections and that there is certain matrix that tabulates whether predictions fall into one
convexity in the resulting mask of four categories:

True positives (TP): P): These refer to the positive
These problems can be reduces by using tuples that were correctly labelledlabel by the
morphological operation such as opening, closing and classifier.
assifier. It is assumed that TP is the number of
filling process. true positives.

True negatives (TN): These are the negative tuples
D. Feature Extraction that were correctly labelleded by the classifier. It is
The foremost features of the Melanoma Skin Lesion assumed that TN is the number of true negatives.
are its Asymmetric Index, Border features, Colo
Colour and
False positive (FP): These are the negative
ne tuples
Diameter. Hence, this system is proposed to extract that were incorrectly labelled
labe as positive. It is
the (9) Geometric Features, (1) Asymmetry Index and assumed that FP is the number of false positives.
(1) diameter of the segmented skin lesion. These
False negative (FN): These are the positive tuples
features are adopted from the segmented image that were mislabelled led as negative. It is assumed
containing only skin lesion, the image blob of the skin that FN is the number of false negatives.
lesion is analyzed to extract
act the 11 features
features.
Accuracy
racy can be calculated by using this equation:
e
Asymmetry Features: Major and Minor Asymmetry
Indices: 
   
   
Accuracy  (1)
Geometric Features:  
 
  

@ IJTSRD | Available Online @ www.ijtsrd.com | Volume – 2 | Issue – 6 | Sep-Oct

Oct 2018 Page: 964
 International Journal of Trend in Scientific Research and Development (IJTSRD) ISSN: 2456-6470
2456
TP  TN

TP  TN  FP  FN

As indicated in Equation (1),, accuracy only measures

the number of correct predictions of the classifier and
ignores the number of incorrect predictions.

Sensitivity, also known as recall,, is computed as the

fraction of true positivess that are correctly identified.
TP Fig.2. Testing result of pre-processing
pre and
Sensitivity = (2)
TP + FN segmentation

Precision, which is computed as the fraction of

retrieved instances that are relevant.
TP
Precision = (3)
TP + FP

Specificity,, computed as the fraction of true negatives

that are correctly identified.
TN
Specificity = (4)
TN + FP
Fig.3. Testing result of segmented image after dilation
and erosion
IV. Test And Results
A. Results of Segmentation Process B. Results of Classification
There are three main phases: namely pre pre-processing, The first stage of the system is to select skin lesion
segmentation and post-processing
processing in segmentation images. This can click original image menu item from
part. The input of the system is dermoscopic image oof main window GUI of the proposed system.
skin lesion as shown in Fig. 2.. After resizing the input
image, this image will convert to the gray scale image After loading the input image, it is needed to segment
in order to get grater separability between the lesion the lesion image by using Otsu's O Thresholding
and background healthy skin. The resultant gray scale Technique. The segmented image obtained from
image has been displayed. Most of the dermoscopic Otsu's thresholding has the advantages of smaller
images have some artifacts such as oil, bubble hair, storage space, fast processing speed and ease in
noise etc. these artifacts are removed ved by applying manipulation,
n, compared with gray level image which
median filter on gray scale image. After this, the usually contains 256 levels.
Otsu’s thresholding method has been applied on the
gray scale intensity image. This gives the desired In feature extraction, the standard features such as
segmented image. Asymmetry Index, Area, Perimeter, Major Axis
Length, Minor Axis length, Circularity Index,
After these steps morphological operations are used to Irregularity Index and diameter are extracted from the
enhance the segmented
ented image. In this research work, segmented
ted test image as shown in Fig.6.Fig. These
morphological opening, dilation, erosion and closing standard features are veryery useful to classify the
operations are used. A morphological closing of the melanoma skin cancer more accurately.
mask is performed
med using a disk of radius 500 pixels
followed by a dilation using a disk of radius 40 pixels. After extracting the features, these features are given
These operations smooth the border of the mask. The as input to the classifier.
er. The classifier produces
experimental results for this post-processing
essing phase are whether the image is Melanoma or not. For the
shown in Fig. 3. Melanoma condition, the classifier output
o is 1 and for
normal skin (not melanoma) the output is 0. By
pressing the classify button, the classification process
is performed and the result
ult is displayed as shown in
Fig. 7 and 8.

@ IJTSRD | Available Online @ www.ijtsrd.com | Volume – 2 | Issue – 6 | Sep-Oct

Oct 2018 Page: 965
 International Journal of Trend in Scientific Research and Development (IJTSRD) ISSN: 2456-6470
2456

Fig.4.
4. Loading Input Image Fig.7.
7. Classification Results of Melanoma

5. After Segmentation
Fig.5. 8. Classification Results of Non-Melanoma
Fig.8. Non

C. Performance Evaluation
To evaluate performance in this system, 40 unknown
images from a test data set are tested. The accuracy of
skin lesion classification system is calculated. Table
1 describes the performance of the proposed system.

TABLE I Performance of The Proposed System

N Im Acc Pre Spe

F T F Sensi
o age TP urac cisi cific
P N N tivity
. Set y on ity

Tes
t 92.5 94.7 90 90.4
1 18 2 19 1
ing % % % 7%
Set
Fig.6. Extracted Features from Testing Image

@ IJTSRD | Available Online @ www.ijtsrd.com | Volume – 2 | Issue – 6 | Sep-Oct

Oct 2018 Page: 966
 International Journal of Trend in Scientific Research and Development (IJTSRD) ISSN: 2456-6470
2456
V. Conclusions References
In this work, the system for melanoma skin cancer 1. Xu L, Jackowski
ckowski M, Goshtasby A, Roseman,
detection system is developed by using MATLAB. “Segmentation
tation of cancer images”, Image and
Image segmentation is the first step in early detection Vision Computing, Volume 17, Issue1, pg. 65-74.
65
of melanoma skin cancer. To analyze skin lesions, it 2. J Abdul Jaleel, Sibi Salim, Aswin. R. B,
is necessary to accurately locate and isolate the
“Computer Aided Detection 01 Skin Cancer”,
lesions. In this thesis work, the Otsu’s method is the International Conference on Circuits, Power and
oldest and simplest one. It has shown the best
Computing Technologies, 2013
segmentation results among the three methods.
Feature extraction is considered as the most critical 3. Arroyo J L G and Zapirain B G,
G “Computer Aided
state–of-the- art skin cancer screening system. In this Detection 01 Skin Cancer”, 2014, Computers in
thesis, the feature extraction is basedd on ABCD
ABCD-rule Computers in biology and .medicine 44 14157
of dermatoscopy. Algorithms for extracting features biology and medicine 44 14157
have been disused. All the features have been 4. Dr. S. Gopinathan, S. Nancy Arokia Rani,
calculated based on Otsu’s segmentation method. “Feature Extraction through Image Processing
Using the Neural Network classifier, melanoma skin Techniques”, International
national Journal of Emerging
cancer diagnosis with a training accuracy of 100% Trends & Technology in Computer Science
and testing accuracy of 93% is achieved. (IJETTCS), Volume 5, Issue
sue 4, July-August
July 2016,
Computational time is around 14 seconds fo for each Web Site:www.ijettcs.org
lesion classification. For illustration, a graphical user
interface has developed in order to facilitate the 5. Uzma Bano Ansari, “Skin Cancer Detection Using
diagnostic task for the dermatologists. Image Processing”, International Research Journal
of Engineering and Technology (IRJET), Volume:
Acknowledgment 04 | Apr-2017
Firstly, the author would like to acknowledge 6. Shivangi jain, Vandana jagtap, Nitin Pise,
particular thanks to Union Minister of the Ministry of “Computer Aided Melanoma Skin Cancer
Science and Education, for permitting to attend the Detection Using Image Processing”, Internal
Master program at Pyay Technology University. Conference on Intelligent Computing
Much gratitude is owed to Dr. Nyaunt Soe, Rector, Communication & Convergence, Vol 48, pp 725- 725
Pyay Technological University, for his kind 740, 2015,
015, ISSN 1877-0509,
1877
permission to carry out this paper. The author iis http://www.sciencedirect.com/science/article/pii/S
.sciencedirect.com/science/article/pii/S
deeply thankful to herr supervisor, Dr. Moe Mon 1877050915007188
Myint, Professor, Department of Electronic
Engineering, Pyay Technological University, fofor her 7. M. Elbaum, “Computer-aided
“Computer melanoma
helpful and for providing guidelines. Moreover, the diagnosis”, Dermatologic clinics, vol. 20, pp. 735-
735
author wishes to express special thanks to Dr. Khin 747, 2002
Thu Zar Win, in, Professor and Head, Department of
Mechatronic Engineering, Pyay Technological
University, Pyay for her kindness and suggestions.
Finally, I would like to thank my parents for
supporting to me.

@ IJTSRD | Available Online @ www.ijtsrd.com | Volume – 2 | Issue – 6 | Sep-Oct

Oct 2018 Page: 967