=Paper=
{{Paper
|id=Vol-2820/SP4HC_paper4
|storemode=property
|title=Decision Support Software for Melanoma Skin Cancer Detection (DECIME)
|pdfUrl=https://ceur-ws.org/Vol-2820/AAI4H-13.pdf
|volume=Vol-2820
|authors=Carlos E. B. Sousa,Leonardo B. G. Nascimento,Cláudio M. S. Medeiros,Mauro Oliveira,Iago A. Trajano
|dblpUrl=https://dblp.org/rec/conf/ecai/SousaNMOT20
}}
==Decision Support Software for Melanoma Skin Cancer Detection (DECIME)==
https://ceur-ws.org/Vol-2820/AAI4H-13.pdf
Decision Support Software for Melanoma Skin
Cancer Detection (DECIME)
Carlos E. B. Sousa and Leonardo B. G. Nascimento and Cláudio. M. S. Medeiros and
Mauro Oliveira1 and Iago A. Trajano2
Abstract. In this paper is proposed a software to decision support [19], Multilayer Perceptron (MLP) [14] and Support Vector Machine
for the detection of melanoma cancer. This approach is proposed, (SVM) [22] to make decision, providing healthcare professionals
as this type of skin cancer is the only one that can metastasize, that with a technological tool and giving the patients the possibility of
is, proliferate to other organs such as lungs, liver, etc. The proposed a faster treatment.
system is elaborated through the acquisition of a set of images with The paper has the following organization: the Section 2 presents
17805 samples, extractors of attributes gray level co-occurrence ma- the related works that helped in the construction of this work. In Sec-
trix (GLCM), Local Binary Pattern (LBP) and Central Moments. For tion 3 it is described the proposed method, that is, the acquisition of
the training and sample classification process, the Single Layer Per- the set of images, elaboration of data sets, parameterization of the
ceptron (SLP), Multilayer Perceptron (MLP) and Support Vector Ma- machine learning algorithms and implementation of the solution in
chine (SVM) classifiers are used. Subsequently, the best-evaluated a Raspberry Pi. The results obtained are presented in Session 4 and,
model is implanted in a Raspberry Pi computer that, together with a finally, Section 5, the final considerations are presented.
webcam and a computer screen, allows the capture and classification
of skin lesions in real-time. In the applied methodology, the authors
obtained the best result of a 93 % accuracy from the use of Central 2 RELATED WORKS
Moments extractor and MLP classifier. In contrast to the state of the Mariam A. Sheha, Mai S. Mabrouk and Amr Sharawy [12] propose
art of this problem, a high level of similarity is found between the an automated method applied to a set of dermoscopic images for the
accuracy rates of the Single Layer Perceptron (SLP) and the Multi- diagnosis of melanoma. The extracted resources are based on Gray
layer Perceptron (MLP), demonstrating a possible resolution of the Level Co-occurrence Matrix (GLCM) and use of the MultiLayer Per-
problem in a linear format. ceptron classifier (MLP) to classify melanoma cancer. The classifier
chosen was used with two different techniques in the training and
1 INTRODUCTION testing process: traditional and automatic. Traditional MLP obtained
a superior performance to automatic MLP. While automatic MLP ob-
Skin cancer is currently presented as the most common illness in the tained a 93.4 % accurac y of the training set and 76 % on the testing
world among fair-skinned people, the non-melanoma type being the set, traditional MLP obtained 100 % accuracy on the training set and
most common kind of skin cancer worldwide. There are two major 92 % on the testing set.
types of skin cancer: melanoma and non-melanoma [18][7]. On another approach, H. Alquran et al [4] propose a method
Melanoma skin cancer is a malign tumor that stems from for detection, presentation, extraction and classification to detect
the uncontrolled proliferation of melanocytes (pigment producing melanoma, using image processing techniques that were applied on
cells)[2][1][11][16]. There are recent records of increase in mortality dermoscopic image samples suspect of melanoma. The classification
rates recurring from melanoma. The costs for melanoma treatment system uses a SVM to classify the lesions. The study’s SVM classi-
can be evaluated in several billions to countries that display a larger fier obtained 92,1% accuracy in the classification, and appears to be a
incidence of the disease. Several preventive strategies have been im- promising approach on the distinction of skin lesions of both benign
plemented in various high risk regions, with different success rates. and malign melanoma.
Over the last four decades, melanoma incidence has increased world- Ansari, U. B. and Sarode, T. [6] propose a system for early de-
wide, the highest of which being in Australia, where there are records tection of skin cancer. The diagnosis methodology uses image pro-
of 40 new cases for every 100.000 citizens per year [13]. cessing techniques associated with a SVM classifier. The dermo-
In this paper, therefore, is proposed a decision support software scopic skin cancer images are obtained and submitted to several pre-
tool that employs techniques from Computer Vision and Artifi- processing techniques for noise removal and image enhancement.
cial Intelligence using the classifiers Single Layer Perceptron (SLP) They are subsequently submitted to image segmentation by thresh-
olding. The authors obtained an accuracy of 95 % with the aforemen-
1 Federal Institute of Education, Science and Technology of Ceará, tioned techniques.
Brazil, email: carlosestevaobs, leonardobrendoti,claudiosa1965, amaurobo-
liveira@gmail.com
2 Federal University of Ceará, Brazil, email: iagoatrajano@gmail.com
Copyright © 2020 for this paper by its authors. Use permitted under Cre- 3 THE PROPOSED METHOD
ative Commons License Attribution 4.0 International (CC BY 4.0). This
volume is published and copyrighted by its editors. Advances in Artificial As previously mentioned, the authors of this work propose a decision
Intelligence for Healthcare, September 4, 2020, Virtual Workshop. support software for melanoma skin cancer detection. That said, as
shown in Fig. 1, the process is divided into 6 steps. The initial 4 3.2 Elaborating the dataset
stages consist of implementing the computational tool, while the last
two stages comprehend practical use of the software. The authors propose the use of the extractors Gray Level Co-
occurrence Matrix (GLCM) [8], Local Binary Pattern (LBP) [17]
and Central Moments. The GLCM describes texture through a set
of characteristics for the occurrences of each level of gray in the pix-
els of the image considering multiple directions [5]. In this way, the
size of the matrix is determined from the distinct number of levels of
pixels contained in the original image [20]. Haralick, K. Shanmugam
and Dinstein [8] proposed a method to which 14 statistical measures
of texture can be obtained from its use, however, only 13 are actually
used, since the latter presents computational instability.
D. Huang et al. [9] state that in recent years, LBP has sparked
a growing interest in image processing and computer vision. This
method efficiently summarizes local image structures, comparing
each pixel with its neighbor pixels. P. Mohanaiah and P. Sathya-
narayana and L. GuruKumar [15] cite that one of the most impor-
tant properties of this operator is computational simplicity, it makes
Figure 1. Steps of Implementation and use for the proposed software. it possible to analyze images in challenging configurations in real-
time.
Moments are the statistical expectation of certain power functions
of a random variable. The central moment is widely used in pattern
As shown on Fig. 1, the implementation of the tool can be divided recognition because of their discrimination power and robustness [3].
into 4 stages: the acquisition of the image set, the elaboration of the The main advantage of this extractor is their invariances to transla-
data set, the use of machine learning techniques and the deployment tions of the object. Therefore they are suited well to describe the
on Raspberry Pi. The methodologies used for each illustrated step shape of the object [10], which can become a strong tool for the elab-
are presented next. oration of a data set that represents well the problem of the detection
of melanoma cancer.
3.1 Acquiring the image set
3.3 Applying machine learning techniques
The information is extracted from an image set named ”dermoscopic
Regarding the artificial intelligence algorithms, the single-layer Per-
pigmented skin lesions from HAM10k”3 . This set contains 8903
ceptron algorithm is implemented to verify the possibility of dealing
samples of melanoma skin cancer images and 8902 samples of non-
with linear problems and, subsequently, the Multilayer Perceptron
melanoma skin cancer images (Fig. 2) , with a total of 17805 sam-
(MLP) and the Support Vector Machine, as they present themselves
ples.
as robust algorithms to obtain better results, in case of a nonlinear
problem.
To verify the efficiency of the proposed methodology, experiments
are carried out with the Single Layer Perceptron (SLP) classifier, us-
ing learning rates of 0.01, 0.05, 0.1 and 0.5, signal function and re-
duction linear or exponential learning rates.
For the MLP classifier, we implemented several learning rates that
vary between 0.1 and 0.5, linear or exponential reduction of these
Figure 2. Example of samples used. rates, logistic function or hyperbolic tangent. The values referring to
the number of neurons in the hidden layer of the MLP are entered
empirically, ranging from 2 to 100 neurons.
For the SVM classifier, 4 types of kernels are used; linear, RBF,
As for the ABCDE rule, professionals use The first five letters of polynomial and sigmoid. The value of C, that is, the penalty param-
the alphabet a guide to help people recognize the warning signs of eter is entered from the following values 0.01, 0.1, 1.0 and 10.
melanoma. They are: Asymmetry (A) - wounds or stains are pre- For both algorithms, the training is conditioned to end when it
sented asymmetrically; Borders (B) - tend to be irregular; Colors (C) reaches an error rate of 10−5 and there are 2000 iterations, in which
- they present different colors in the same wound; D (Diameter) - the weights with the best accuracy are chosen as the ones that best
equal to or greater than 6mm; and Evolution (E) - whether it evolves represent the results of the training.
in shape, color or elevation [21]. As mentioned, the authors used 17.805 images in total, which are
In this sense, based on the aforementioned rule, the authors pro- divided into 8.903 samples of the class ”melanoma” and 8.902, of
pose the use of attribute techniques based on textures and shapes, as ”non-melanoma”. In this sense, 7.122 samples selected at random
they are similar to the way of detecting the problem in a practical for each type of class were used to compose the training/tests set.
way. The techniques used are presented next. The rest of the data, that is, 1.781 samples of the class ”melanoma”
and 1.780 samples of the class ”non-melanoma”, make up the test
3 The data set is made available by Alexander Scarlat through the link: set.
https://www.kaggle.com/drscarlat/melanoma For all classifiers, the K-Fold cross-validation technique is used,
where a value of K equals 10. The application of this method consists 4 RESULTS OBTAINED
of dividing the total set of training data into K subsets of the same
size (Fig. 3). Thus, each subset is used for validation, while the rest In the Table 1 shows the best results obtained in this study. This solu-
of the set is applied to estimate the parameters. tion is obtained with the Central Moments extractor and Multilayer
Perceptron classifier. It should be noted that, in order to obtain the
best results, the authors use various combinations between their at-
tributes generated by the extractors. Thus, were obtained the best
results (Table 1) by applying the following Central Moments: mu11,
mu21, mu12, mu30, mu03.
Table 1. Results obtained from the extraction of attributes with Central
Moments and classification with Multilayer Perceptron.
Stage N. Acc Prec. Rec. F1
Mel 1.00 1.00 0.88 0.94
Train
Non-mel. 0.89 0.90 1.00 0.94
Mel 0.88 0.99 0.87 0.93
Test
Non-mel. 0.99 0.89 0.99 0.94
Although the Multilayer Perceptron classifier offers a better solu-
tion from the data obtained by the Central Moments extractor, the
Figure 3. Training and testing processes. authors perceive similar results from the same attributes used with
the Single Layer Perceptron classifier. This perception demonstrates
that the problem of detecting melanoma cancer can be linearly sepa-
rable, making it an easily resolvable problem. Table 2 illustrates the
As shown in Fig. 3, the data set is divided into 2 parts. The first best results obtained from the classifier SLP.
is divided into 10 subsets that are used in the training and validation
process. Finally, the second part of the set, unknown by the classifier, Table 2. Results obtained from the extraction of attributes with Central
is used for the test process. Moments and classification with Single Layer Perceptron.
Stage N. Acc Prec. Rec. F1
Mel 0.97 0.97 0.89 0.93
3.4 Deploying the software in Raspberry Pi 3 B+ Train
Non-mel. 0.90 0.90 0.98 0.94
computer Test
Mel 0.89 0.97 0.89 0.93
Non-mel. 0.98 0.90 0.98 0.93
In order to make this application usable to healthcare professionals
as well as to attain low acquisition costs, the computer program (Fig.
4) was deployed to a Raspberry Pi 3B+ computer. It is important to emphasize that by using the Central Moments
extractor, one can divide the data almost linearly. Therefore, the de-
cision limits for both classifiers have similarities and simplicity in
their structures.
5 CONCLUSIONS
In view of the literature involved, this work was differentiated by
using less robust techniques and achieving similar or even superior
results. Providing a lower computational cost in the early detection
Figure 4. Deploying the program to a Raspberry Pi 3 B+ computer with a of melanoma cancer.
webcam. Although most Artificial Intelligence models require powerful
processing and extensive memory resources, there are still methods
that, when applied correctly, produce satisfactory results. Therefore,
After installing the classification software on the raspberry com- the exact application of an attribute extractor related to the binary
puter, we can use the webcam to obtain images of the skin for analy- detection of melanoma skin cancer produces a good solution to the
sis, and the computer screen for verifying the suggested results. problem without requiring much processing power or memory. To
that extent, the use of attribute extractors suitable for a problem can
simplify its classification. In this scenario, a Perceptron can be a sig-
3.5 Confirming the detection with a doctor and nificant resource for achieving satisfactory results.
generating results It should be noted that when deploying the software on the rasp-
berry pi board, a simple, fast and portable solution is created, which
As previously mentioned, the authors propose a decision support enables support the decision to health care professionals regarding
software. Thus, the results generated by it must be analyzed by a melanoma skin cancer detection. The software also helps improve
doctor for proper conclusions. accessibility to patients who live in regions of difficult access, since
the device is portable enough for professionals to provide home care
when the patient is unable to reach the care unit.
REFERENCES
[1] Z Abdel-Malek, V B Swope, I Suzuki, C Akcali, M D Harriger, S T
Boyce, K Urabe, and V J Hearing, ‘Mitogenic and melanogenic stim-
ulation of normal human melanocytes by melanotropic peptides.’, Pro-
ceedings of the National Academy of Sciences, 92(5), 1789–1793,
(1995).
[2] Tada A Im S Akcali C. Abdel-Malek Z, Suzuki I, The melanocortin-1
receptor and human pigmentation, Acad Sci., Ann N Y, 2nd edn., 1999.
[3] Sundos A Hameed Al azawi, ‘Eyes recognition system using central
moment features’, Engineering and Technology Journal, 29(7), 1400–
1407, (2011).
[4] H. Alquran, I. A. Qasmieh, A. M. Alqudah, S. Alhammouri,
E. Alawneh, A. Abughazaleh, and F. Hasayen, ‘The melanoma skin
cancer detection and classification using support vector machine’, in
2017 IEEE Jordan Conference on Applied Electrical Engineering and
Computing Technologies (AEECT), pp. 1–5, (2017).
[5] S Annadurai, Fundamentals of digital image processing, Pearson Edu-
cation India, 2007.
[6] Uzma Bano Ansari and Tanuja Sarode, ‘Skin cancer detection using
image processing’, Int Res J Eng Technol, 4(4), 2875–2881, (2017).
[7] Skin Cancer Foundation. Melanoma warning signs - what you need to
know about early detection, 20-.
[8] Robert M Haralick, Karthikeyan Shanmugam, and Its’ Hak Dinstein,
‘Textural features for image classification’, IEEE Transactions on sys-
tems, man, and cybernetics, (6), 610–621, (1973).
[9] Di Huang, Caifeng Shan, Mohsen Ardabilian, Yunhong Wang, and
Liming Chen, ‘Local binary patterns and its application to facial image
analysis: a survey’, IEEE Transactions on Systems, Man, and Cyber-
netics, Part C (Applications and Reviews), 41(6), 765–781, (2011).
[10] Yaser Daanial Khan, Nabeel Sabir Khan, Shoaib Farooq, Adnan Abid,
Sher Afzal Khan, Farooq Ahmad, and M Khalid Mahmood, ‘An effi-
cient algorithm for recognition of human actions’, The Scientific World
Journal, 2014, (2014).
[11] Aaron B. Lerner and Joseph S. McGuire, ‘Melanocyte-stimulating hor-
mone and adrenocorticotrophic hormone’, New England Journal of
Medicine, 270(11), 539–546, (1964). PMID: 14091524.
[12] Mai Mabrouk, Mariam Sheha, and Amr Sharawi, ‘Automatic detection
of melanoma skin cancer using texture analysis’, International Journal
of Computer Applications, 42, 22–26, (03 2012).
[13] Vishal Madan, John T Lear, and Rolf-Markus Szeimies, ‘Non-
melanoma skin cancer’, The Lancet, 375(9715), 673 – 685, (2010).
[14] Warren S. McCulloch and Walter Pitts, ‘A logical calculus of the ideas
immanent in nervous activity’, The bulletin of mathematical biophysics,
5(4), 115–133, (Dec 1943).
[15] P Mohanaiah, P Sathyanarayana, and L GuruKumar, ‘Image texture
feature extraction using glcm approach’, International journal of sci-
entific and research publications, 3(5), 1, (2013).
[16] K. Nouri, Skin Cancer, MGraw Hill professional, McGraw-Hill Educa-
tion, 2007.
[17] Timo Ojala, Matti Pietikainen, and Topi Maenpaa, ‘Multiresolution
gray-scale and rotation invariant texture classification with local binary
patterns’, IEEE Transactions on pattern analysis and machine intelli-
gence, 24(7), 971–987, (2002).
[18] D.S. Rigel, R. Friedman, J.K. Robinson, M.I. Ross, C.J. Cockerell,
H. Lim, E. Stockfleth, and J.M. Kirkwood, Cancer of the Skin E-Book:
Expert Consult, Elsevier Health Sciences, 2011.
[19] F. Rosenblatt, ‘The perceptron: a theory of statistical separability in
cognitive systems (project para)’, (1958).
[20] William Robson Schwartz and Hélio Pedrini, ‘Método para
classificação de imagens baseada em matrizes de coocorrência
utilizando caracterı́sticas de textura’, III Colóquio Brasileiro de
Ciências Geodésicas, CuritibaPR, Brasil, 1, (2003).
[21] Marina Tsatmali, Janis Ancans, and Anthony J Thody, ‘Melanocyte
function and its control by melanocortin peptides’, Journal of Histo-
chemistry & Cytochemistry, 50(2), 125–133, (2002).
[22] V. N. Vapnik, ‘The nature of statistical learning theory’, Spring - Verlag
New York, (1995).