10th International Conference on Electrical and Computer Engineering

20-22 December, 2018, Dhaka, Bangladesh 185

Feature Extraction from Dermoscopy Images for an

Effective Diagnosis of Melanoma Skin Cancer
Sharmin Majumder* and Muhammad Ahsan Ullah
Department of Electrical and Electronic Engineering,
Chittagong University of Engineering and Technology, Chittagong-4349, Bangladesh
*
[email protected]

Abstract—The aim of this paper is to extract some distinct notched. “C” represents Color: the color is not uniform all
geometric features from dermoscopy images to classify benign over the lesion. “D” is for Diameter: the lesion is larger than 6
and malignant melanomas. To avoid skin biopsy which is an millimeters, although melanoma lesions may sometimes be
invasive technique, diagnosis of melanoma skin cancer from smaller than this. According to Patwardhan et al. [4], this
dermoscopy images was developed. It is a very challenging task
ABCD rule diagnoses melanoma only for thin melanocytic
due to some reasons. Firstly, high degree of intraclass variation
exists among melanoma images while low interclass variation is lesions and it has 59–88% accuracy in detection of malignant
found between melanoma and non-melanoma images. Secondly, melanoma, but for precise diagnosis, biopsy is essential.
benign and malignant melanoma images are visually similar to a About 75-84% diagnostic accuracy was determined even
great extent. And finally noises like hair are always present in when the expert dermatologists use the dermoscopy images
skin images which make difficult to analyse the images. In this [5]. There are mainly four proposed rules and methods such as
work, we used fundamental ABCD rule to detect malignant seven-point checklist method [6], ABCD rule [2], the Menzies
melanoma and benign lesion based on quantitative measures. In method [7] and pattern analysis [8] to determine the
our proposed technique, we extracted a new feature which is the melanoma skin cancer. In 2015, Marín et al. [9] noticed that
difference between maximum and minimum Feret diameters of
the ABCD rule became inefficient while detecting micro-
the best fit ellipse to skin lesion. This discriminative feature alone
classified the melanomas with 86.5% accuracy. In our approach, melanomas. This evaluation determines a specificity of 91%
we applied the feature extraction block containing all parameters and sensitivity of 43%. Amarathunga et al. [10] proposed a
to 200 images and the overall accuracy of 98% was achieved to system which enabled patient to detect skin diseases via
detect malignant and benign melanoma from the images. A online and the system provided therapeutic advice also. They
Back-propagation Neural Network (BNN) model was developed employed some data mining classification algorithms such as
and eventually used as a classifier in this proposed method. AdaBoost, BayesNet, J48, MLP and Naive Bayes to classify
three types of skin diseases and achieved 80-85% accuracy for
Index Terms— Skin Cancer, Melanoma, Lesion, Digital Image melanoma detection. Jain et al. in 2015 used ABCD features
Processing, Artificial Neural Network, Segmentation, Feature [11]. Krishna et al. [12] proposed several clustering
Extraction. techniques for segmentation process and extracted features by
using ABCD. In this paper, we developed five effective
I. INTRODUCTION
features to classify malignant and benign melanoma. These
Mainly three types of skin cancers, Basal-Cell features are Asymmetry score (AS), Border Irregularity (B),
Carcinoma (BCC), Squamous Cell Carcinoma (SCC) Color variegation (C), Diameter (D1) and Difference between
and melanoma are found frequently. Among them melanoma maximum and minimum Feret diameters of the best fit ellipse
is the deadliest form of skin cancer although it is not the most to the lesion (D2). In our proposed technique, we achieved
common. It is responsible only for 4% of all cancers occurred accuracy of 98% with 95% sensitivity and 98.8% specificity.
in human skin while it is accountable for 75% of deaths
caused by skin cancers [1]. This cancer is primarily caused II. PROPOSED METHODOLOGY
due to severe exposure of skin to the sunlight. It is prime need The system flow diagram of the proposed methodology for
to detect melanoma in the early stage because if it is treated in automated diagnosis of malignant and benign melanoma skin
the initial state, it is almost curable. But if the cancer is not lesion is presented in Fig. 1. Potential skin lesion image which
detected and treated from its early stage, it spreads adjacent is perceived to be cancerous is applied as input to the system.
parts of the body rigorously when it becomes tenacious to Afterwards image pre-processing and segmentation are
treat and ultimately causes death. The American Cancer performed to the original image to adjust the image quality
Society estimates that in 2018, about 91,270 new melanomas and to separate lesion from surrounding skin background.
are suspected to be diagnosed (about 55,150 in male and Here Otsu threshold algorithm has been used as segmentation
36,120 in female) in the United States. Almost 9,320 people technique. The segmented lesion is carried out through feature
are presumed to die of melanoma (about 5,990 male and 3,330 extraction block which contains some analysis to determine
female) [2]. Australia and New Zealand have the highest rates the essential features. A BNN (Back-propagation Neural
of melanoma in the world [3]. The American Cancer Society Network) model is trained with the extracted features from
[2] provides the ABCD rule which is fundamental method for 200 images. Finally, the classification accuracy, sensitivity
the dermatologists and steerage of self-examination for the and specificity of the proposed model have been computed to
patient for Malignant Melanoma. “A” stands for evaluate the system.
Asymmetry: one half of the lesion does not match another. “B”
is for Border: the verges are irregular, uneven, blur, ragged or

978-1-5386-7482-6/18/$31.00 ©2018 IEEE

 186

from the surrounding normal skin. After lesion segmentation,

other smaller blobs may be present in the image. Some
morphological operations such as “hole filling” find the
biggest blob in the segmented image which is only the skin
lesion. Fig. 2(a) represents an original image collected from
PH2 database [13] and Fig. 2(b) represents the corresponding
segmented image.

a) Source Image b) Segmented Image

Fig. 2 Source Image and segmented image of a skin image.

C. Feature extraction
At first the image blob is analysed to extract area, centroid,
perimeter, orientation. Area (A) specifies the sum of pixels in
Fig. 1 Work Flow Diagram for the proposed technique of melanoma skin the segmented lesion. Perimeter (P) characterizes the distance
cancer detection between each adjoining pair of pixels around the border of the
contour. Centroid represents centre of mass of the lesion.
A. Input Image Datasets Orientation defines the angle between the x-axis and the major
Dermoscopic images for our research work were taken axis of the best fit ellipse to the object.
from PH2 database which was collected at Pedro Hispano
Hospital [13]. The PH2 is a dermoscopic image database of 1) Asymmetry: Asymmetry is the most important parameter to
automatic computer-based diagnosis system for dermoscopy distinguish malignant from benign melanoma. Asymmetry
images (ADDI) project which contains skin lesions of describes that one half of the melanoma lesion does not match
different sizes, shapes and colors. A joint research another half in size and shape. A melanoma lesion is usually
collaboration of the University of Porto and the Dermatology considered asymmetric across both axes (horizontal and
Service of Hospital Pedro Hispano, Matosinhos, Portugal vertical). We translate and rotate the skin lesion by its
developed the PH2 database [13]. The dermoscopic images orientation angle to align the major axis with the x-axis of the
were acquired in the same conditions through Tuebinger mole image coordinate system as shown in Fig. 3. The half region
analyzer system which used a magnification of 20× optical of rotated lesion (T) along x-axis is masked with black and
zoom. The contained images are 8-bit RGB color images with flipped across x axis as shown in Fig. 4(a). These two half
a resolution of 768×560. We used total 200 dermoscopic regions are added (TX1) and the non-overlapping region (X1)
images, containing 160 benign melanomas and 40 malignant between T and TX1 is determined from (1). This procedure is
melanomas. also applied for the other half of the lesion across x-axis (Fig.
4(b)). The procedure stated above is repeated for the y axis
B. Image Pre-Processing also which is shown in Fig. 4(c) and Fig. 4(d).
Before applying the main processing, the input image is
passed through a noise removal system. Here, we apply
averaging filter to the true-color (RGB) input image. The
source skin image may be collected in any non-uniform
illumination. In order to compensate this condition, some
image pre-processing techniques are needed. Here, the Image
pre-processing means Image enhancement which includes
image resizing and contrast adjustment. After the pre-
processing, the next step is to segment the lesion from the
surrounding skin image. The main purpose of the
segmentation step is to split an image into regions that are
homogeneous in terms of pixel intensity or specific features.
Segmentation enhances the edges of the lesion boundary while
subduing the gradients in both inside and outside regions of
the lesion. Because of color variation in different datasets, T

color images are converted into the intensity mode. Lesion Fig. 3 Binary Image rotated to determine the asymmetry with respect to its
segmentation is done by automatic thresholding and then axes.
masking operation in gray scale image. First, Otsu’s automatic
The following mathematical formula are used to estimate
threshold is applied. In Otsu's method, threshold is determined
the non-overlapping regions.
by trial and error which minimizes the intra-class variance and
maximizes the interclass variance [14]. The main prerequisite X 1 = T ⊕ TX 1 , X 2 = T ⊕ TX 2 , X = X 1 + X 2 (1)
for feature extraction block is to separate the lesion accurately Y1 = T ⊕ TY1 , Y2 = T ⊕ TY2 , Y = Y1 + Y2 (2)
 187

Asymmetry Score across horizontal axis (AS1) and vertical color is 5% of total pixels in the lesion, then the color is
axis (AS2) are determined by the (3) assumed to be present. And the score of color is determined as
ΔX ΔY the number of colors estimated in the lesion. This score ranges
AS1 = , AS 2 = (3) from 1 to 6. Fig. 5 is characterized by two colors: dark-brown
ΔT ΔT
Where, ΔX, ΔY = sum of pixels in non-overlapping regions and light-brown. Hence the color score is 2.
X and Y respectively. ΔT= total pixels in lesion. 4) Lesion Diameter: Since, a melanoma mole is
considered to grow larger than a common mole, it is one of
the most important parameters. A melanoma lesion generally
has a diameter greater than 6mm [2]. In this paper we first
calculate the average diameter (DImage) of a skin lesion using
(5) [18]. The acquired diameter is in number of pixels. Then
to obtain the actual value of diameter in mm, we use (6).
4A (5)
TX1 X1 TX2 X2 DImage =
π
DImage
D Actual = × 0.2645 (6)
S
where, S represents the magnification factor of original image.
X1
In our work, S = 20 and 0.2645 is the conversion factor from
pixel to mm.
TY1 Y1 TY2 Y2
5) Difference between Maximum and Minimum Feret’s
Diameters (D2): In this work, we proposed a new feature
Fig. 4 Non-overlapping regions of original image and folded image across x- which is the difference between maximum Feret diameter (M)
axis (a, b) and y-axis (c, d). and minimum Feret diameter (m) of the best fit ellipse to skin
2) Border Irregularity: Melanoma lesion is considered to lesion. For evaluating this factor, at first best fit ellipse to
have ragged, uneven, blur, and irregular border. To determine lesion has been constructed which is shown in the top left
the border irregularity (B), here circularity index is used as image of Fig. 4. We built the ellipse having same normalized
given in (4). This score ranges from 0 to 1. When the border second central moments as that of the lesion. Then the M and
of a lesion is uneven, ragged, and irregular, this score m of the ellipse are measured and difference between them
approaches to zero [15]. (D2) is calculated. The acquired value is in number of pixels.
4πA From the analysis it is found that, difference between Feret
B= 2 (4) diameters increases when the lesion grows to be melanoma.
P
For benign lesion, D2 is of much lower value.
3) Color Variegation: Normal lesion consists of a uniform
D. Neural Network for Classification
color [2]. Usually they are characterized by brown, black etc.
But one important sign of a lesion to be melanoma is color After pre-processing and feature extraction have been done,
variation in the lesion. Malignant melanomas contain three or features extracted before are used to train the backpropagation
more types of colors. Even five or six colors may be present in neural network (BNN). The output of NN is compared with
almost 40% of melanoma lesions [16]. In [15], authors desired output and if it does not match then an error signal is
converted the true color image to HSV color space. Then they generated.
calculated color variance of the skin lesion and concluded a
range of variance for melanoma colors. Wilson et al. [17]
obtained a range of colors using histogram after converting of
HSV. Regarding to color variegation, six different colors are
considered in our work according to PH2 database [13]. The
colors which might be existing in malignant melanoma lesion
are white, light brown, dark brown, red, blue gray and black.

Fig. 6 Structure of Neural Network

To minimize the error, weights are being adjusted and this

process continues until the error becomes zero. The network
consists of an input layer with an activation function, at least
one hidden layer and output layer. We used sequential order
Fig. 5 Color Variation score
weight/bias training ‘trains’ as the training function and
In this paper we determined number of pixel ranges need to hidden layer with hundred neurons as shown in Fig. 6. The
be present for each red, green and blue channel to form above dataset is split into three categories: 70% for training data, 15%
mentioned six colors. If the number of pixels of an individual for validation and 15% for testing.
 188

III. RESULTS AND DISCUSSION developed some significant and suitable features of melanoma
TABLE I represents three sample images from the database dermoscopy images which are very easy and fast to calculate.
and their extracted features. The classification accuracy (A) of Then we used fully supervised back propagation NN to
skin images is assessed with sensitivity (Sn) and specificity classify malignant and benign melanoma. In our approach, the
(Sp). We calculated these performance indicators according to weights are the same for all images and this model has been
(7, 8, 9). The experimental results are given in TABLE II. proved to be accurate for all types of images. Since the
TP proposed method shows an accuracy of considerable extent,
Sn = (7) this will be an efficient and precise method for diagnosis of
TP + FN melanoma skin cancer.
TN
Sp = (8) REFERENCES
TN + FP
[1] C. Lu, M. Mahmood, N. Jha, and M. Mandal, “Automated
TP + TN segmentation of the melanocytes in skin histopathological images,”
A= (9)
TP + FN + TN + FP IEEE J. Biomed. Heal. Informatics, vol. 17, no. 2, pp. 284–296,
2013.
Here, TP, FN, TN, FP specify True Positive, False Negative, [2] American Cancer Society, “Melanoma Skin Cancer.” [Online].
True Negative, and False Positive, respectively. Available: https://www.cancer.org/cancer/melanoma-skin-cancer.
[Accessed: 25-Jun-2018].
TABLE I [3] P. Boyle and B. Levin, “World Cancer Report,” 2014.
EXPERIMENTAL RESULTS FOR THREE SAMPLE IMAGES [4] S. V. Patwardhan, S. Dai, and A. P. Dhawan, “Multi-spectral image
analysis and classification of melanoma using fuzzy membership
Original Segmented Features Output based partitions,” Comput. Med. Imaging Graph., vol. 29, no. 4, pp.
Image Image 287–296, 2005.
AS1: 0.13 [5] G. Argenziano et al., “Dermoscopy of pigmented skin lesions:
AS2: 0.13 Results of a consensus meeting via the internet,” J. Am. Acad.
B: 0.86 Dermatol., vol. 48, no. 5, pp. 679–693, 2003.
Benign [6] G. Argenziano, G. Fabbrocini, P. Carli, V. De Giorgi, E. Sammarco,
C: 2
D1: 4 and M. Delfino, “Epiluminescence microscopy for the diagnosis of
doubtful melanocytic skin lesions: Comparison of the ABCD rule of
D2: 0.96
dermatoscopy and a new 7-point checklist based on pattern analysis,”
AS1: 0.16 Arch. Dermatol., vol. 134, no. 12, pp. 1563–1570, 1998.
AS2: 0.18 [7] S. W. Menzies, C. Ingvar, K. A. Crotty, and W. H. McCarthy,
B: 0.86 “Frequency and morphologic characteristics of invasive melanomas
Benign
C: 1 lacking specific surface microscopic features,” Arch. Dermatol., vol.
D1: 4.3 132, no. 10, pp. 1178–1182, 1996.
D2: 0.46 [8] H. Pehamberger, A. Steiner, and K. Wolff, “In vivo
epiluminescence microscopy of pigmented skin lesions. I. Pattern
AS1: 0.24
analysis of pigmented skin lesions,” J. Am. Acad. Dermatol., vol. 17,
AS2: 0.18 no. 4, pp. 571–583, 1987.
B: 0.5 [9] C. Marín, G. H. Alférez, J. Córdova, and V. González, “Detection
Malignant
C: 4 of melanoma through image recognition and artificial neural
D1: 8.6 networks,” in IFMBE Proceedings, 2015, vol. 51, pp. 832–835.
D2: 1.5 [10] A. A. L. C. Amarathunga, E. P. W. C. Ellawala, G. N. Abeysekara,
and C. R. J. Amalraj, “Expert System For Diagnosis Of Skin
TABLE II Diseases,” Int. J. Sci. Technol. Res., vol. 4, no. 01, pp. 174–178,
RESULTS OF THE PROPOSED METHOD 2015.
[11] S. Jain, V. Jagtap, and N. Pise, “Computer aided melanoma skin
Actual Actual Sensitivity Specificity Accuracy cancer detection using image processing,” in Procedia Computer
Benign Malignant (Sn) (Sp) (A) Science, 2015, vol. 48, no. C, pp. 736–741.
Predicted as [12] M. Chaithanya Krishna, S. Ranganayakulu, and D. P. Venkatesan,
TN =158 FP =2
Benign “Skin Cancer Detection and Feature Extraction through Clustering
95% 98.8% 98% Technique,” Int. J. Innov. Res. Comput. Commun. Eng., vol. 4, no. 3,
Predicted as
FN =2 TP =38 pp. 3736–3742, 2016.
Malignant
[13] T. Mendonca, P. M. Ferreira, J. S. Marques, A. R. S. Marcal, and J.
TABLE III Rozeira, “PH2- A dermoscopic image database for research and
COMPARATIVE STUDY WITH PREVIOUS WORKS benchmarking,” in Proceedings of the Annual International
Conference of the IEEE Engineering in Medicine and Biology
Method Year Accuracy Society, EMBS, 2013, pp. 5437–5440.
[16] 2016 93.24% [14] N. Otsu, “A Threshold Selection Method from Gray-Level
Histograms,” IEEE Trans. Syst. Man. Cybern., vol. 9, no. 1, pp. 62–
[18] 2017 95.45%
66, 1979.
Proposed Method 2018 98% [15] O. Shridharmurumkar and P. Gumaste, “Feature Extraction for Skin
Cancer Lesion Detection,” Int. J. Sci. Eng. Technol. Res., vol. 4, no.
Performance of our method is also evaluated through 5, pp. 1645–1650, 2015.
[16] J. Jaworek-korjakowska and J. Jaworek-korjakowska, “Computer-
comparison based on overall classification accuracy with the
Aided Diagnosis of Micro-Malignant Melanoma Lesions Applying
methods proposed in [16] & [18]. The comparative study is Support Vector Machines,” vol. 2016, no. June, 2016.
shown in TABLE III. A close observation of the table reveals [17] G. D. Wilson F. Cueva, F. Muñoz, G. Vásquez. and Escuela,
that our proposed technique gives more promising result “Detection of skin cancer ‘Melanoma’ through Computer Vision,”
2017.
comparing others’ procedures.
[18] Nidhal K. EL Abbani and Zahraa, “Detection and Analysis of Skin
Cancer from Skin Lesions,” Ijaer, vol. 12, no. 19, pp. 9046–9052,
IV. CONCLUSIONS 2017.
The number of melanoma cancer patient have been rising
for the last 30 years, hence effective and fast cancer detection
method in the early stage is of great importance. Here we