Hindawi

BioMed Research International

Volume 2020, Article ID 5345923, 14 pages
https://doi.org/10.1155/2020/5345923

Research Article
New Auxiliary Function with Properties in Nonsmooth Global
Optimization for Melanoma Skin Cancer Segmentation

Idris A. Masoud Abdulhamid ,1 Ahmet Sahiner ,1 and Javad Rahebi 2

1
Department of Mathematics, Suleyman Demirel University, Isparta, Turkey
2
Department of Electrical and Computer Engineering, Altinbas University, Turkey

Correspondence should be addressed to Javad Rahebi; [email protected]

Received 28 June 2019; Revised 26 January 2020; Accepted 14 February 2020; Published 14 April 2020

Academic Editor: Nasimul Noman

Copyright © 2020 Idris A. Masoud Abdulhamid et al. This is an open access article distributed under the Creative Commons
Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work
is properly cited.

In this paper, an algorithm is introduced to solve the global optimization problem for melanoma skin cancer segmentation. The
algorithm is based on the smoothing of an auxiliary function that is constructed using a known local minimizer and smoothed
by utilising Bezier curves. This function achieves all filled function properties. The proposed optimization method is applied to
find the threshold values in melanoma skin cancer images. The proposed algorithm is implemented on PH2, ISBI2016
challenge, and ISBI 2017 challenge datasets for melanoma segmentation. The results show that the proposed algorithm exhibits
high accuracy, sensitivity, and specificity compared with other methods.

1. Introduction optimization problems. Thus, from a technical and scien-

tific point of view, it is becoming increasingly important
Skin cancers are the most widely recognised type of growths to study and develop methods that can solve this class of
in humans. They are a type of destructive disease that affects difficult mathematical problems. Currently, numerous new
skin [1]. Most skin growths are reparable during the early theoretical and practical methods have been reported to
stages. Therefore, the early diagnosis of skin tumours can search for the global optimum. The one-dimensional opti-
save patients. Nowadays, computers and intelligent handheld mization method is one of the most effective approaches
devices are common and therefore can help diagnose mela- for this purpose. This method is based on selecting direc-
noma earlier. Computer-Aided Diagnosis (CAD) tools can tions randomly. The group of one-dimensional algorithms
be connected to such devices to create a smart system that is referred to as line search methods [5]. They define an
helps dermatologists recognize the melanoma. Moreover, algorithm in which the direction of search is determined
Codella et al. [2] have shown in recent research that some randomly at each iteration. This class of optimization
of these CAD systems have better performance than average methods is used as a part of descent techniques, which
human expert agreements. Traditional melanoma detection depend on objective function derivatives. The filled func-
CAD programs generally consist of three main components: tion method was produced by Ge (1990) [6]. It is produced
the segmentation of the lesion, the extraction of functions, with respect to passing from the current local minimizer to
and the classification of features [3]. Yu et al. [4] showed that the lower one until the global minimizer is determined.
although melanoma patients can only be classified by means Numerous classes of filled functions have been introduced
of features extracted using deep study models, the diagnostic by several authors (see, e.g., [6–12]). Bezier curves were
performance is significantly improved by incorporating seg- introduced to solve the nonsmoothness of curves. They
mented lesions. were developed by Pierre Bezier in 1962 for the styling of
Numerous problems in the fields of engineering, eco- motorcar bodies [13]. Currently, Bezier curves are broadly
nomics, and natural sciences can be represented as global used in computer graphics and animation. They are
2 BioMed Research International

normally used for curve and surface design. In this study, a or tan. Melanomas are considered to be the most dangerous
new auxiliary function with properties in nonsmooth global of skin growths. Even though Merkel cell carcinoma is lethal
optimization is developed. This auxiliary function achieves in most cases, melanoma generally causes more deaths than
all filled function properties, and it is used for finding the several other types of skin disease. According to the Ameri-
threshold value in melanoma skin cancer images. Even can Cancer Society, approximately 76,380 (46,870 male and
though the proposed algorithm has demonstrated powerful 29,510 female) new instances of melanoma were reported
in both local and global searches and can be widely used in in 2016, among which 10,130 cases were fatalities (6750 male
various optimization problems, it has some limitations and 3380 female). The frequency of melanoma has been ris-
including parameters adjustment, as it is a delicate task ing globally each year. Numerous lives can be saved if mela-
since it directly affects the efficiency of the auxiliary function. nomas can be identified at the earliest stages when they are
This paper contributes to the development of melanoma effectively treatable. Various examination methods based on
detection using new auxiliary function with properties in different advances are being widely developed for the early
nonsmooth global optimization algorithm. In this study, identification of melanoma.
the melanoma in the dermoscopic images was detected by Jain and Pise [24] proposed methods of detecting mel-
three steps. The red channel on the colored dermoscopy anoma skin cancer using computer systems and image
image is selected in the preprocessing step. A 2D median- processing. Numerous studies have demonstrated that
size 25 ∗ 25 filter and morphological filtering based on the computer vision can play a vital role in medical image
Gaussian kernel have been used for smoothing. The follow- diagnosis. The contribution to the framework was the skin
ing step was to determine the optimum melanoma segmen- sore figure. After that, by applying novel figure preparing
tation threshold value using the proposed algorithm. procedures, it is examined to finish near the skin growth.
Eventually, the estimated optimum threshold value has Jaleel et al. [25] detected skin cancer using computer sys-
been used as the thresholding method used in the Otsu tems. The biopsy technique is commonly used for skin disease
[14]. In order to demonstrate the effectiveness of this study, recognition. In biopsy, skin is evacuated or scratched and sam-
the proposed method has been evaluated on three public ples undergo extensive testing. Computer-based skin growth
dermoscopic image databases: PH2 [15], ISBI2016 chal- identification is beneficial to patients who can distinguish a
lenge [16], and ISBI 2017 challenge [17]. This paper is skin malignancy without visiting a healing centre or without
organized as follows. In Section 2, related works are viewed. the assistance of a specialist. Computer-based identification
In Section 3, the new method is presented, preceded by pre- utilises imaging strategies and artificial intelligence. The dis-
liminaries and assumptions related to global optimization. tinctive phases of identification include the accumulation of
In Section 4, the melanoma detection using global optimi- dermoscopic figures, search of the figures for expelled hairs
zation algorithm is proposed. Evaluation results with com- and noise, fragmentation of the figures by utilising the maxi-
parisons are reported in Section 5. Finally, the conclusion mum entropy threshold, extraction of highlights using a gray
of the study is remarked in Section 6. level coevent matrix, and arrangement by employing an artifi-
cial neural network. A backpropagation neural network is uti-
2. Related Works lised for ordering. It categorises a given informational index as
cancerous or noncancerous.
The conventional method for determining skin malignancy Esteva et al. [26] proposed techniques for the segmenta-
is the biopsy technique [3, 18]. In this technique, skin is tion of different types of skin carcinoma. Profound convolu-
scratched or evacuated, and samples are collected for tional neural networks (CNNs) indicated the potential for
research and testing. Computer-based skin disease location general and specific factor assignments crosswise over multi-
is beneficial to patients because patients can distinguish skin ple fine-grained protest classes.
tumours without visiting a healing centre or without the Ramlakhan and Shang [27] proposed a classification sys-
assistance of a specialist [3]. tem for cancerous skin lesions. In their technique, a model of
Xie et al. [19] proposed novel convolutional neural net- a figure was constructed by a robotised melanoma acknowl-
work for skin lesion segmentation. In their method, they gen- edgement framework using Android cell phones. The frame-
erated high-resolution feature maps to preserve spatial work comprised of three noteworthy segments, i.e., figure
details. For the enhancement of features, the special and division, computation, and characterisation. It was intended
channel-wise mechanism was adopted. to be run on a cell phone with a camera or on a tablet com-
Hwang and Celebi [20] solved surface in skin figures and puter. Hoshyar et al. [28] investigated the automated early
applied mathematical approaches such as the gray stage detection of skin malignancy. Dermatology imaging researchers
coprevalence matrix. They reported that surface investigation assumed that finding skin melanomas can be automated
may exactly detect the boundary with a smooth surface, and based on certain physical features and shading information
such surface study is the segmentation of dermatological that are typical of the characterisations of skin tumours.
images [21, 22]. Mishra and Celebi [23] reported the detec- Goyal and Jain [29] reported the computer-based automated
tion of skin abnormalities (particularly melanomas) using detection of melanoma skin malignancy. Melanoma skin
image processing methods and machine learning. Melanoma tumours occur when shade-producing cells (melanocytes)
is typically considered as a dull-raised injury; these tumours grow uncontrollably, and these tumours result in pain.
develop from colored cells. A few melanomas lose color, hav- Karargyris et al. [30] proposed a new application for
ing no or almost dull shade, and they can appear pink, white, image processing to identify skin malignancy using a
BioMed Research International 3

smartphone. Xu et al. [31] proposed a technique for the anal- programmed skin malignancy-finding framework that con-
ysis and classification of melanocytic skin cancer. The pro- solidates diverse textural and shading highlights was pro-
posed technique contains four essential modules. In the posed. New textural and shading highlights were utilised
first module, a multiassurance framework parcels the epider- for effective and exact discovery.
mis and dermis zones. Next, an epidermis examination is
performed, in which the nuclear morphologies and spatial 3. Preliminaries of Global Optimization
spreads of the epidermis features are investigated.
Li and Shen [32] proposed a method of analysing skin This section provides definitions and assumptions. A general
lesions using a deep learning network. Two deep learning global optimization problem is indicated as follows:
techniques were proposed to address three principle aspects
in the field of skin injury figure manipulation, i.e., sore min f ðxÞ, ð1Þ
x∈Λ
division, sore dermoscopic highlight extraction, and sore
characterisation. A deep learning system composed of two where Λ ⊂ ℝn is the feasible domain of x that is specified by
completely coevolutionary residual networks (FCRNs) is pro- constraints, and x = ðx1 , ⋯ ⋯ , xn ÞT .
posed to produce segment results and rough configuration Problem (1) is smooth if function f is continuously dif-
results at the same time. A measurement unit for the lesion ferentiable; otherwise, problem (1) is nonsmooth.
index (LICU) was designed to simplify coarse cluster tests by
computing a warm separation chart. For dermoscopic high- Definition 1 (see [36]). The point x∗ ∈ Λ is said to be a global
light extraction, a straight CNN was suggested. Various injury minimizer of f if f ðx∗ Þ ≤ f ðxÞ for all x ∈ Λ.
rates were over 1000 × 700 pixels, requiring high-estimation
costs. A deep learning system requires to rescale of injury esti- Definition 2 (see [8]). The basin, B∗ , of the function, f ðxÞ, at
mates. Moreover, resizing figures can misrepresent the skin an isolated local minimizer, x∗k , is a connected domain con-
injury. In the first case, a sore figure’s middle region has been taining x∗k , in which the steepest descent trajectory of f ðxÞ
edited and then comparatively resized to a lower determina- converges to x∗k from any initial point in B∗ .
tion. The median area value has been set at 0.8 of the figure
height and thus modified with respect to the figure emphasis.
Definition 3 (see [13]). A Bezier curve is characterised by a set
Since the exact probability maps of various skin soreness
of control points, Cn , and it is defined as follows:
groups provide pathologists with valuable data, the LICU has
been suggested to refine the conceived gross skin sore result n
maps derived from FCRNs. Z ðt Þ = 〠 r j,n ðt ÞC j , 0 ≤ t ≤ 1, ð2Þ
Dorj et al. [33] proposed a method for skin cancer classi- j=0
fication using a deep CNN. The focus of the proposed strat-
egy was the assignment of characterising skin tumours by where
utilising an ECOC SVM and a profound CNN. A number !
of figures contained noise such as different organs and appa- n
ratuses. These figures were edited to decrease noise for r j,n ðt Þ = t j ð1 − t Þn−j , ð3Þ
obtaining better results. A current and preprepared AlexNet j
CNN was utilised for removing highlights. An ECOC SVM
classifier was used for skin growth. denotes the Bernstein basis polynomials of degree n.
Cueva et al. [34] proposed a method for skin cancer detec-
tion using computer systems. In their method, figure handling The following assumptions are satisfied in the rest of
was created to produce the asymmetry, border, color, and this paper:
diameter of melanoma by utilising neural systems to group var-
ious types of moles. Subsequently, this calculation was created (A1). The search should ideally have decent directions;
after an examination of 200 figures produced an execution of this implies that
97.51%. The early identification of skin tumours increases the
probability of a cure, such as those found in the cutting-edge d Tk ∇f ðxk Þ < 0: ð4Þ
stages. In this manner, the death rate of this condition may
be reduced. Additionally, late examinations have demonstrated (A2). The search directions should be gradient related, so
that the estimations of the execution on the arrangement of that
melanoma by a dermatologist are in the range of 75 to 84%.
Alfed and Khelifi [35] proposed a technique for detecting kdk k ≥ qk∇f ðxk Þk, ð5Þ
the types of skin cancers from dermoscopic images. Numer-
ous mechanised methods have been proposed to determine where q > 0 is a constant.
and arrange infections to have agreeable skin disease location
execution. Despite this, reducing the false discovery rate is (A3). The determination of δk should include one-
difficult and time consuming because false positives trigger dimensional minimisation. This ensures that
alerts and require mediation by a specialist pathologist for
facilitating examination and screening. In this technique, a f ðxk+1 Þ < f ðxk Þ: ð6Þ
4 BioMed Research International

3.1. New Global Optimization Method. The proposed aux- By utilising multiplication with a piecewise function,
iliary function is constructed based on the best local ψΛ1 ðxÞ, function (7) can be rewritten as
minimizer of f ðxÞ found so far and the elimination
function as γðx, x∗k Þ = f ðx∗k Þ − ½ f ðx∗k Þ − f ðxÞψΛ1 ðxÞ, ð8Þ

γðx, x∗k Þ = min f f ðxÞ, f ðx∗k Þg: ð7Þ where piecewise function ψΛ1 ðxÞ is defined by
(
1, x ∈ Λ1 ,
The typical feature of this function is to remove ΨΛ1 ðxÞ = ð9Þ
local minimizers, higher than the previously found min- 0, otherwise,
imizer, and keep the original, f(x), function unchanged
in a region in which the function values are lower than and includes all possible cases of function γðx, x∗k Þ according
the best value of the algorithm. In other words, it has to the values of x and set Λ1 = fx ∈ Λ : f ðx∗k Þ > f ðxÞg: All
the following properties: terms of function (8) are smooth except for those related to
piecewise function ψΛ1 ðxÞ. Hence, it is sufficient to smooth
(a) If f ðx∗k Þ ≤ f ðxÞ ⟹ γðx, x∗k Þ = f ðx∗k Þ, for all x ∈ Λ function ψΛ1 ðxÞ to ensure that function (8) is smoothed. By
utilising the Bezier curves given by Definition 3, the functions
(b) If f ðx∗k Þ > f ðxÞ ⟹ γðx, x∗k Þ = f ðxÞ, for all x ∈ Λ y1 and y2 can be defined as follows:

   1/2   1/2 
2b2 − 2b1 + 2 ðb2 − 1Þ b21 τ − 2b1 τ + b2 − 1 − 2b1 b2 − 2b1 τ + b21 b2 + b21 τ − 2b1 ðb2 − 1Þ b21 τ − 2b1 τ + b2 − 1 +2
y1 = ,
ðb 1 − 2 Þ 2
   1/2 2
b1 b2 − b2 + b2 b21 τ − 2b1 τ + b2
y2 =   ,
b2 ðb1 − 2Þ2
ð10Þ

( )
where τ = f ðxÞ − f ðx∗k Þ and b1 > 0, 0 < b2 < 1. These func- b2 − b22
tions can be used to obtain the smoothed form of ψΛ1 ðxÞ, 0 ≤ ~γðx, x∗k , b1 , b2 Þ − γðx, x∗k Þ ≤ max b22 , , ð13Þ
b1
which can be written as

8 for all x ∈ Λ.
>
> 0, τ > b2 ,
>
>
>
> b2 ≥ τ > 0, Proof. From the definitions of ~γðx, x∗k , b1 , b2 Þ and γðx, x∗k Þ,
>
> y2 ,
>
> we have
>
< −b2
~ Λ ðτ, b1 , b2 Þ = y1 ,
ψ 0≥τ> , ð11Þ
> b1
> ~γðx, x∗k , b1 , b2 Þ − γðx, x∗k Þ = ð f ðxÞ − f ðx∗k ÞÞ
1
>
>
>
> −b2  
> 1,
> τ≤ :  ψ ~ Λ ðτ, b1 , b2 Þ − ψΛ ðxÞ :
>
>
>
: b1 1 1

ð14Þ

Thus, the smoothed form of function (8) can be written According to the states of τ and parameters b1 and b2 ,
as follows: we consider the following four cases:

Case 1. If τ > b2 , this gives

γðx, x∗k , b1 , b2 Þ = f ðx∗k Þ − ½ f ðx∗k Þ − f ðxÞψ
~ ~ Λ ðτ, b1 , b2 Þ: ð12Þ
1

~γðx, x∗k , b1 , b2 Þ − γðx, x∗k Þ = 0, ð15Þ

Theorem 4. Suppose x∗k
is the local minimizer of f , and
parameters b1 and b2 are defined as above, then we have for x ∈ Λ:
BioMed Research International 5

Case 2. If b2 ≥ τ > 0, we have Therefore, we have

γðx, x∗k , b1 , b2 Þ − γðx, x∗k Þ ≤ b22 ,

~ ð16Þ ~γðx, x∗k , b1 , b2 , aÞ < ~γðx∗k , x∗k , b1 , b2 , aÞ: ð22Þ

for x ∈ Λ:
Thus, x∗k is a local maximiser of ~γðx, x∗k , b1 , b2 , aÞ.
Case 3. If 0 ≥ τ>−b2 /b1 , we have
Theorem 6. Suppose x∗k is a local minimizer of f , then ~γ
ðx, x∗k , b1 , b2 , aÞ has no stationary point for x ∈ Λ2 , where
b2 − b22
γðx, x∗k , b1 , b2 Þ − γðx, x∗k Þ ≤
~ , ð17Þ Λ2 = fx ∈ Λ ∣ f ðxÞ ≥ f ðx∗k Þ, x ≠ x∗k g.
b1
Proof. In case τ = f ðxÞ − f ðx∗k Þ > b2 , we have
for x ∈ Λ.
 
Case 4. If τ≤−b2 /b1 , we have for ~γðx, x∗k , b1 , b2 , aÞ = f ðx∗k Þ + aζ kx − x∗k k2 : ð23Þ

γðx, x∗k , b1 , b2 Þ − γðx, x∗k Þ = 0,

~ ð18Þ For any x satisfying f ðxÞ ≥ f ðx∗k Þ, we have

for x ∈ Λ.  
∇~γðx, x∗k , b1 , b2 , aÞ = a∇ζ kx − x∗k k2 : ð24Þ
As described above, the local removal process loses
significant information and contains several removed local However, ka∇ζðkx − x∗k k2 Þk > 0 for any x ∈ Λ2 , i.e., ~γ
minimizers, which are so hard to handle for the algorithm. ðx, x∗k , b1 , b2 , aÞ does not have a stationary point at x ∈ Λ2 .
Improper implementation also leads to additional complica-
tions in addressing the global optimization problem. Develop- Theorem 7. Suppose x∗k is a local minimizer point of f but not
ing an appropriate method for looking for the best solutions global and f has a lower minimizer than x∗k , then ~γðx, x∗ , b1 ,
that have been identified so far to find better solutions or basins b2 , aÞ has a stationary point in Λ1 = fx ∈ Λ ∣ f ðxÞ < f ðx∗k Þg if
is important. An escape function, ζ, is therefore provided for the
a = jaj ≤ L/T, where k∇ðaζðkx − x∗k k2 ÞÞk ≤ jajT and k∇f k ≤ L.
treatment of removed local minimizers. This feature is based on
the best, x∗k , solution that has been found so far, as follows:
Proof. Selecting parameters b1 and b2 to be sufficiently
small, our smoothed function, ~γðx, x∗ , b1 , b2 , aÞ, can be
γðx, x∗k , b1 , b2 , aÞ =
~ f ðx∗k Þ − ½ f ðx∗k Þ −
~ Λ ðτ, b1 , b2 Þ
f ðxÞψ
  1 obtained as
∗ 2
+ aζ kx − xk k ,  
~γðx, x∗ , b1 , b2 , aÞ = f ðxÞ + aζ kx − x∗k k2 ð25Þ
ð19Þ

where a is a real-value constant. ζ is an escape function that has the in the most part of the lower basin. As the norm of the
form ð1/ð1 + kx − x∗k k2 ÞÞ and satisfies the following properties: gradient of the function,
  1
ζðτÞ > 0, ζ kx − x∗k k2 = , ð26Þ
1 + kx − x∗k k2
ζ ′ ðτÞ < 0, ð20Þ
limτ→∞ ζðτÞ = 0:
is bounded, there exists a number T > 0 such that
The fundamental properties of auxiliary function (19) can be     
 
∇ aζ kx − x∗k k
2
demonstrated by multiple theorems.  ≤ jajT: ð27Þ

Theorem 5. Suppose x∗k is a local minimizer of f and ~γðx,

x∗k , b1 , b2 , aÞ is defined by (19), then point x∗k is a local max- Hence, the following inequality,
imiser of ~ γðx, x∗k , b1 , b2 , aÞ.     
 
∇ aζ kx − x∗k k
2
 ≤ jajT ≤ k∇f k ≤ L, ð28Þ
Proof. As x∗k is a local minimizer of function f , there
exists κ > 0. χ = Nðx∗k , κÞ is a neighbourhood of x∗k , such
that f ðxÞ ≥ f ðx∗k Þ for any x ∈ χ. When x ≠ x∗k , then enables f to increase faster than escape function ζ decreases,
providing a stationary point in Λ1 , which implies
 
~γðx, x∗k , b1 , b2 , aÞ f ðx∗k Þ + a/1 + kx − x∗k k2 L
= < 1: ð21Þ jaj ≤ : ð29Þ
~γðx∗k , x∗k , b1 , b2 , aÞ f ðx∗k Þ + a T
6 BioMed Research International

(A) Initialisation step:

1. Select parameters b1 and b2 ; select and adjust a > 0 according to Theorem 4.
2. Select initial point x0 ∈ Λ.
3. Generate direction d k , k = 1, ⋯, η.
4. Set ε = 10−3 .
(B) Main step:
1. Construct a one-dimensional function for the first direction, dk :
F dk ðδÞ = f ðx0 + δdk Þ:
2. Starting from any arbitrary initial point δ0 , find δik , which is a local minimizer of F dk , and select ϑ = −1.
3. Construct the auxiliary function, ~γðδ, δ∗k , b1 , b2 , aÞ, at δi:k
4. Using δ0 = δk + ϑε, find the minimizer, δγ , of ~γðδ, δ∗k , b1 , b2 , aÞ at δik .
5. If δγ ∈ Λ, go to (6); otherwise, go to (8).
6. Minimise F dk again starting from δγ to find δi+1 k lower than δk , and go to (7).
i

7. If δk ∈ Λ, put δk = δk and go to (1).

i+1 i i+1

8. If ϑ = 1, stop and put δik = δ∗k ; otherwise, take ϑ = 1 and go to (4).

9. Use xk = x0 + δ∗k dk to find xk .
10. Find x∗ of f ðxÞ using xk as the initial point.
11. If k < η, set k = k + 1 and create a new search direction, dk+1 , and then, go to (1); otherwise, go to (7).
12. Select the best global minimizer of function f ðxÞ

Algorithm

The idea of the algorithm can be described in the follow- Step 4. After median filtering, the histogram of the median
ing three important steps: image is calculated. It provides information about the inten-
sity distribution of the median image along with the location
(a) The first step is to reduce the objective function into a of the distribution.
one-dimensional function in each search direction,
dk . The one-dimensional functions are found using
F dk ðδÞ = f ðx0 + δdk Þ as a function of δ Step 5. The proposed global optimization algorithm is
applied to the histogram of the median image using Otsu’s
(b) The second step is to construct an auxiliary function, method. The weighted sum of the variance of the histogram
γðδ, δ∗k , b1 , b2 , aÞ, at δik . Point δik is any arbitrary local
~ is considered as a fitness function.
minimum of F dk ðδÞ, and it is used as an initial point
to search for the global minimizer in direction d k Step 6. Finally, the global optimization method provides the
(c) The third step is to repeat the above phases for all threshold value for detecting melanoma. This threshold
search directions to obtain the local minimizers of f value is compared with the median filtering image. If a
ðxÞ. The lowest among these minimizers is the point of the image is larger than the threshold value, then
required global minimizer of f ðxÞ. the point is changed to white; otherwise, the point is chan-
ged to black.
4. Melanoma Detection Using Global
Optimization Algorithm Step 7. After the image is converted to black and white, an
image overlay function is applied to mask the corner of the
In this section, the method of global-optimization-based image. After the masking, the final detected image is gener-
melanoma detection is explained step by step. ated and compared with the ground truth image.
Sensitivity, specificity, positive predicted value, negative
Step 1. The original image is fetched from a database. predicted value, accuracy, and computation time are ana-
lysed for effective comparison.
Step 2. The ground truth image is fetched from the database.
The flowchart for global-optimization-based melanoma
Step 3. In current melanoma skin cancer segmentation, the detection is shown in Figure 1. The steps of global optimiza-
original grayscale image is preprocessed by employing two tion for skin cancer segmentation are shown as a graphical
transformations. One is the conversion of the grayscale abstract in Figure 2. In this study, for the validation of the pro-
image to RGB, and the other is median filtering conversion. posed optimization process, the PH2 dataset has been used.
This conversion eliminates hue and saturation but maintains Every coefficient was evaluated by a competent dermatologist
luminance. Median filtering reduces noise such as salt-and- for the numbering factors of analytic performance, the manual
pepper noise. partition of lesion areas, and the dermoscopic norm [37].
BioMed Research International 7

Proposed optimization
Preprocessing algorithm Skin cancer segment

Load original image Initialize the parameters of Compare all pixel of

and ground truth global optimization the image with the
method best global value

Calculate the Get the segmented

Convert original characteristic function image and remove
RGB image to ∼
𝜓(x) the noises from the
grayscale image segmented result

Calculate the auxiliary

function Compare the
∼ segmented result
Apply median filter 𝛾(x,x⁎,b1,b2)
with 20 ⨯ 20 mask with the ground
truth result

Calculate:
𝛿k = 𝛿0 – 𝜈𝜀
and ⁎ Display segmented
Calculate the area of skin cancer
histogram of image xk = x0 – 𝛿0dk
and find
data with 256 bins performance
metrics
Find the maximum
variance value
of the image

Detection of skin tumors

Figure 1: Flowchart for the segmentation process using the proposed global optimization method.

5. Evaluation Results Figure 3 shows the processed melanoma images used

to prove the effectiveness of the proposed algorithm. The
The International Skin Imaging Collaboration recommends effectiveness of the proposed method is proved by comparing
several metrics for performance evaluation. These metrics it with a newly developed method, i.e., the ASLM method∗
indicate the performance of each algorithm and method that (details available at 10.1016/j.com-pmedimag.2016.05.002).
categorises pixels correctly. Such metrics show the output of Six melanoma images are randomly selected, and detection
each algorithm and process that correctly categorises pixels. is carried out using the proposed method and ASLM
This analysis measures precision (Acc (%)), specificity (Spe method. The images are shown in Figure 4. The final
(%)), positive prediction (PPV), negative predictive value images obtained by the proposed method are more accu-
(NPV), and sensitivity (Sen (%)). The following equations rate than those obtained by the ASLM method. Table 1
are used for the quantitative analysis of the results of all the shows the comparison of the performance parameters of
techniques applied by the various parameters: the global optimization algorithm with JSEG, SRM, KPP,
K-means, Otsu, Level Set, and ASLM for 200 image detec-
Ntp + Ntn
AC = , tions. Based on these results, the proposed method pro-
Ntp + N f p + Ntn + N f n duces superior parameters, with a specificity of 0.9928
Ntp and an accuracy of 0.9011.
JA = , The sensitivity of the proposed method higher than
Ntp + N f n + N f p
that of the other methods, except the ALSM method.
2 × Ntp
DI = , ð30Þ The results obtained using the proposed global optimiza-
2 × Ntp + N f n + N f p tion method on the PH2 database images of malignant
Ntp lesions (melanomas) show IMD088 (blue-whitish veil,
SE = , streaks, and regression areas), IMD284, IMD405, and
Ntp + N f p
IMD419 (blue-whitish veil) for the second, third, and
Ntn
SP = : fourth rows, IMD424 (blue-whitish veil and streaks) for
Ntn + N f p the fifth row, and IMD425 (blue-whitish veil and
8 BioMed Research International

Observation Create Model Extract Global Build Algorithm

optimization

Test
Refine
Simulation

Preprocessing of the image

Load original Conversion of Apply median Calculate

image and color image filtering to histogram for
ground truth into gray gray image median
image image filtered image

Nonsmoothed global optimization

Compare threshold value with

Apply proposed method median filtered image to
to find threshold value convert black and white image

Melanoma detection stage and performance analysis

Create masking Performance Final detected

image for corner analysis of the image compared
using overlay skin cancer with ground
function detected image truth image

Parameters (accuracy (Acc (%)), specificity (Sep (%)), sensitivity (Sen (%)), (NPV) and (PPV)

Detection of skin tumors

Figure 2: Steps of global optimization method for skin cancer segmentation.

regression areas) for the sixth row. The computational and the values of the parameters in the algorithm are
speeds for all processing steps are provided in Table 2. taken as follows:
All performance parameters are calculated utilising a
stratified cross validation method, in which the PH2 data- b1 = 0:1, b2 = 0:4, a = 1, ε = 0:003, υ = −1,
base is divided into three subgroups, each with approxi-
G = 100, M = 15, ubðthe upper boundÞ ð31Þ
mately 40 unhealthy moles and 160 healthy moles for
skin cancer lesions. Table 3 shows a few performance = 10, lbðthe lower boundÞ = 0:
parameters of the six processed images selected randomly
from the PH2 database. A recognised statistical assess- From the ISBI 2016 and ISBI 2017 databases, the
ment parameters are applied to experimentally compare most complicated images (i.e., images with dermoscopes,
the implementation of the proposed algorithm with exist- bubbles, hair, and multiple colors) have been used in
ing state-of-the-art segmentation algorithms. The pro- their original dimensions. These databases contain 8-bit
posed algorithm is applied to 200 benign and melanoma dermoscopic RGB images from 540 ∗ 722 to 4499 ∗ 6748
medical images obtained from the PH2 database [38], pixels, with different image dimensions. In addition, both
BioMed Research International 9

Image Origina Ground truth Median Detected

number imagel image filtering image image

(IMD437)

(IMD436)

(IMD434)

(IMD431)

(IMD423)

(IMD207)

(IMD146)

(IMD101)

(IMD063)

(IMD020)

Figure 3: Ten sets of images for melanoma detection using global optimization algorithm.

databases included the original images coupled with the Figure 6 also compares, in terms of Sen, Spe, Acc, Dic,
boundaries of lesion segmentation, which have been recorded and Jac, the proposed approach and other common methods
by professional dermatologists. employed by the ISBI 2017 database.
Figure 5 lists the segmentation results, where the mea- According to uneven skin patches, such as freckles, der-
surement of the comparative methods was taken from the moscopic image repositories could have several small objects.
original papers. Deeper learning methods achieve better Such small objects can be screened out by using the
results than the method in [39], and with the proposed median filter. In some cases, all impurities are not removed
method, the highest Jac and Dic values are obtained. There- and the filter has an impact on algorithm work and results.
fore, the proposed method outperforms comparable methods Figure 7 shows images that fail to the segment. The color
and segments of skin lesions effectively. contrast of skin lesion and underlying skin in rows 2 and 3 is
10 BioMed Research International

Proposed
Image Original Ground truth method
number image image ASLM result

(IMD088)

(IMD284)

(IMD405)

(IMD419)

(IMD424)

(IMD425)

Figure 4: Comparison of proposed method and ASLM method.

Table 1: Performance parameter results for 200 image detections Table 3: Comparison of performance parameters for the five
from the PH2 database. processed images selected randomly from the PH2 database.

Method Sen Spe Acc Image no. PPV NPV Computation time (sec)
Otsu (MATLAB 2014a) 0.5221 0.7064 0.6518 1 0.901184 0.999942 3.312
SRM [49] 0.6518 08757 0.6766 2 1 0.988564 3.407
KPP (MATLAB 2014a) 0.4147 0.9581 0.7815 3 0.984021 1 3.489
Level set [50] 0.7188 0.8003 0.7842 4 0.970134 0.997304 3.579
K-means (open CV2.4) 0.7291 0.8430 0.8249 5 0.924157 0.998954 3.357
JSEG [51] 0.7108 0.9714 0.8947 6 0.9327 0.9898 3.217
ASLM 0.8024 0.9722 0.8966
Global optimization algorithm 0.8892 0.9933 0.932
100
90
80
70
60
Table 2: Total time consumed by the proposed method for six 50
40
images selected randomly from 200 images in the PH2 database. 30
20
10
Proposed method 0
Total time consuming (sec) SE SP ACC DIC JAC
(global optimization) Yu et al. [4] 91.1 95.7 94.9 87.7 82.9
Codella et al. [2] 69.3 83.6 80.7
Image processing 19:25 ± 2:02 Menegola et al. [40] 47.6 88.1 79.2

ð18:27 ± 12:13Þ × 10−3

Vasconcelos et al. [41] 74.6 84.5 82.5
Classification Oliveira et al. [42] 91.8 96.7 27.7
Global optimization algorithm 93.42 97.31 95.24 88.51 84.61

Figure 5: Comparison of earlier studies with the current method for

ISBI 2016 database [2, 4, 40–42].
BioMed Research International 11

100
90
80
70
60
50
40
30
20
10
0
SE SP Accuracy DIC JAC
Yuan et al. [43] 82.5 97.5 93.4 84.9 76.5
Bi et al. [44] 42.7 96.3 85.8 84.4 76
Li & Shen [32] 82 97.8 93.2 84.7 76.2
Al-masni et al. [45] 85.4 96.69 94.03 87.2 77.11
Guo et al. [38] 97.5 88.8 95.3 90.38 80.32
Jahanifar et al. [46] 81 98.1 93 83.9 74.9
Tschandl et al. [47] 85.3 77
Bi et al. [48] 96.7 91.4 86.2 85.7 77.7
Global optimization algorithm 98.64 98.52 97.61 92.31 84.26

Figure 6: Comparison of earlier studies with the current method for ISBI 2017 database [32, 38, 43–48].

(a) (b) (c) (d)

Figure 7: A variety of segmentation failure cases: Row (1-3) PH2, ISBI 2016, and ISBI 2017 datasets; Columns (a) original image; (b)
manual segmentation image (ground truth); (c) proposed method segmentation image; and (d) result of proposed method (green) and
ground truth (red).

identical. In this case, the optimal threshold value is difficult rithm is suitable for finding and segmenting melanoma
to determine because of color differences between the skin skin cancer.
lesions and the background with no noticeable pigmentation
in the lesions. If the contrast between the skin lesion and 6. Conclusion
the skin around is not sufficient enough, the gray threshold
algorithm of global optimization takes in a large part of the In this paper, a new method is proposed for melanoma
skin around the image and stretches to the edge of the skin cancer segmentation. The method is based on the
image. It also causes a segmentation failure as the mask global optimization technique, and it uses an auxiliary
reaches the image’s edge. Considering that the primary skin function for performing directional search via Bezier
lesion is an adjacent area, segmented images should first curves. The proposed algorithm is a fully automatic
involve only one adjacent context with no isolated elements detection technique that does not require a training
or troughs. The first area (including the image) can be of phase. Moreover, its computation time for detecting mel-
any size and possibly bordering the image. This is why anoma images is extremely short. The proposed method
the image failed in row 1. is verified experimentally using three public dermoscopic
The comparison of state-of-the-art methods and the image databases. The performance of the proposed algo-
global optimization algorithm is provided in Table 4. If rithm is compared with the existing melanoma segmenta-
sensitivity is considered as the most significant compo- tion techniques in terms of sensitivity, specificity, and
nent, which implies that every patient with skin cancer accuracy. Evaluation test results show that our proposed
is detected, the global optimization algorithm has the segmentation method outperforms the other conventional
highest performance. Thus, the global optimization algo- state-of-the-art segmentation algorithms, and its efficiency
12 BioMed Research International

Table 4: Comparison of the proposed algorithm with state-of-the-art methods.

Number of cases
References Benign Melanoma Sen (%) Spe (%) Acc (%)
D’Amico and Stanganelli [52] 927 50 96.41 87.16 91.78
Tanaka et al. [53] 181 70 90.00 98.30 94.00
Maglogiannis and Kosmopoulosb [54] 14 20 90.00 93 94.00
Marques et al. [55] 146 17 94.10 77.40 85.50
Olszewska and Semantic [56] 24 24 100 66.66 83.33
Zagrouba and Barhoumi [57] 160 40 71.30 93.50 82.00
Aljanabi et al. [58] 160 40 95.50 98.40 96.02
Proposed global optimization algorithm 160 40 95.93 98.99 96.28

is comparable to the new approaches based on deep neu- [5] V. Gardeux, R. Chelouah, P. Siarry, and F. Glover, “EM323:
ral networks. The proposed method is suitable for mela- a line search based algorithm for solving high-dimensional
noma segmentation in skin cancer diagnosis. continuous non-linear optimization problems,” Soft Com-
puting, vol. 15, no. 11, pp. 2275–2285, 2011.
[6] G. Ren-Pu, “The theory of filled function method for finding
Data Availability global minimizers of nonlinearly constrained minimization
problems,” Journal of Computational Mathematics, vol. 5,
The PH2 database means the Pedro Hispano Hospital that
no. 1, pp. 1–9, 1987.
deals with dermatology by dermoscopic images in a Tubin-
gen Mole Analyzer system. This data set is free and open [7] F. Wei, Y. Wang, and H. Lin, “A new filled function method
with two parameters for global optimization,” Journal of Opti-
source and all people they can use these data set. Also we
mization Theory and Applications, vol. 163, no. 2, pp. 510–527,
used these data set to our proposed method. 2014.
[8] Y. Zhang, Y. Xu, and L. Zhang, “A filled function method
Ethical Approval applied to nonsmooth constrained global optimization,” Jour-
nal of Computational and Applied Mathematics, vol. 232, no. 2,
This article does not contain any studies with human partic- pp. 415–426, 2009.
ipants or animals performed by any of the authors. [9] Y. M. Liang, L. S. Zhang, M. M. Li, and B. S. Han, “A filled
function method for global optimization,” Journal of Compu-
Conflicts of Interest tational and Applied Mathematics, vol. 205, no. 1, pp. 16–31,
2007.
The authors declare that there is no conflict of interest [10] A. Sahiner, N. Yilmaz, and O. Demirozer, “Mathematical
regarding the publication of this paper. modeling and an application of the filled function method in
entomology,” International Journal of Pest Management,
vol. 60, no. 3, pp. 232–237, 2014.
Acknowledgments
[11] H. Liu, Y. Wang, X. Z. Gao, C. Dang, and Z. Meng, “A param-
The authors would like to thank the Dermatology Service of eter free filled function method for global optimization,”
Pedro Hispano Hospital, Matosinhos, Portugal for making Pacific Journal of Optimization, vol. 14, no. 4, pp. 567–580,
2018.
their database available.
[12] A. Sahiner and S. A. Ibrahem, “A new global optimization
technique by auxiliary function method in a directional
References search,” Optimization Letters, vol. 13, no. 2, pp. 309–323,
2019.
[1] B. Intraocular, Melanoma Treatment (PDQ): Health Profes-
[13] P. Bezier, “Mathematical and practical possibilities of UNI-
sional Version, PDQ Cancer Information Summaries, 2015.
SURF,” Computer Aided Geometric Design, pp. 127–152, 1974.
[2] N. C. F. Codella, Q. B. Nguyen, S. Pankanti et al., “Deep learn-
ing ensembles for melanoma recognition in dermoscopy [14] N. Otsu, “A threshold selection method from gray-level histo-
images,” IBM Journal of Research and Development, vol. 61, grams,” IEEE Transactions on Systems, Man, and Cybernetics,
no. 4/5, pp. 5:1–5:15, 2017. vol. 9, no. 1, pp. 62–66, 1979.
[3] N. Z. Tajeddin and B. M. Asl, “A general algorithm for auto- [15] N. Dey, V. Rajinikanth, A. Ashour, and J. M. Tavares, “Social
matic lesion segmentation in dermoscopy images,” in 2016 group optimization supported segmentation and evaluation
23rd Iranian Conference on Biomedical Engineering and 2016 of skin melanoma images,” Symmetry, vol. 10, no. 2, p. 51,
1st International Iranian Conference on Biomedical Engineer- 2018.
ing (ICBME), pp. 134–139, Tehran, Iran, November 2016. [16] A. Casari, J. Chester, and G. Pellacani, “Actinic keratosis and
[4] L. Yu, H. Chen, Q. Dou, J. Qin, and P.-A. Heng, “Automated noninvasive diagnostic techniques: an update,” Biomedicines,
melanoma recognition in dermoscopy images via very deep vol. 6, no. 1, p. 8, 2018.
residual networks,” IEEE Transactions on Medical Imaging, [17] ISIC, ISIC 2017: Skin Lesion Analysis towards Melanoma Detec-
vol. 36, no. 4, pp. 994–1004, 2017. tion, 2017, June 2018, https://challenge.kitware.com/
BioMed Research International 13

#challenge/n/ISIC 2017: Skin Lesion Analysis Towards Mela- [33] U. O. Dorj, K. K. Lee, J. Y. Choi, and M. Lee, “The skin cancer
noma Detection. classification using deep convolutional neural network,” Mul-
[18] B. Stewart and C. P. Wild, World Cancer Report, WHO, 2014. timedia Tools and Applications, vol. 77, no. 8, pp. 9909–9924,
[19] F. Xie, J. Yang, J. Liu, Z. Jiang, Y. Zheng, and Y. Wang, “Skin 2018.
lesion segmentation using high-resolution convolutional neu- [34] W. F. Cueva, F. Muñoz, G. Vásquez, and G. Delgado, “Detec-
ral network,” Computer Methods and Programs in Biomedi- tion of skin cancer melanoma through computer vision,” in
cine, vol. 186, p. 105241, 2020. 2017 IEEE XXIV International Conference on Electronics, Elec-
[20] S. Hwang and M. E. Celebi, “Texture segmentation of dermo- trical Engineering and Computing (INTERCON), Cusco, Peru,
scopy images using Gabor filters and G-means clustering,” in August 2017.
IPCV 2010: Proceedings of the 2010 International Conference [35] N. Alfed and F. Khelifi, “Bagged textural and color features for
on Image Processing, Computer Vision, & Pattern Recognition, melanoma skin cancer detection in dermoscopic and standard
pp. 882–886, Las Vegas NV, USA, 2015. images,” Expert Systems with Applications, vol. 90, pp. 101–
[21] T. Yao, Z. Wang, Z. Xie, J. Gao, and D. D. Feng, “A multiview 110, 2017.
joint sparse representation with discriminative dictionary for [36] E. K. Chong and S. H. Zak, An Introduction to Optimization
melanoma detection,” in 2016 International Conference on (Vol. 76), John Wiley & Sons, 2013.
Digital Image Computing: Techniques and Applications [37] T. Mendona, P. M. Ferreira, J. S. Marques, A. R. Marcal, and
(DICTA), pp. 1–6, Gold Coast, QLD, Australia, November J. Rozeira, “PH2-a dermoscopic image database for research
2016. and benchmarking,” in 2013 35th Annual International Con-
[22] R. Moussa, F. Gerges, C. Salem, R. Akiki, O. Falou, and ference of the IEEE Engineering in Medicine and Biology Society
D. Azar, “Computer-aided detection of melanoma using geo- (EMBC), pp. 5437–5440, Osaka, Japan, July 2013.
metric features,” in 2016 3rd Middle East Conference on Bio- [38] Y. Guo, A. Ashour, and F. Smarandache, “A novel skin lesion
medical Engineering (MECBME), pp. 125–128, Beirut, detection approach using neutrosophic clustering and adap-
Lebanon, October 2016. tive region growing in dermoscopy images,” Symmetry,
[23] N. K. Mishra and M. E. Celebi, “An overview of melanoma vol. 10, no. 4, p. 119, 2018.
detection in dermoscopy images using image processing and [39] F. Xie, H. Fan, Y. Li, Z. Jiang, R. Meng, and A. Bovik, “Mela-
machine learning,” 2016, http://arxiv.org/abs/1601.07843. noma classification on dermoscopy images using a neural net-
[24] S. Jain, V. jagtap, and N. Pise, “Computer aided melanoma work ensemble model,” IEEE Transactions on Medical
skin cancer detection using image processing,” Procedia Com- Imaging, vol. 36, no. 3, pp. 849–858, 2017.
puter Science, vol. 48, pp. 735–740, 2015. [40] A. Menegola, M. Fornaciali, R. Pires, F. V. Bittencourt, S. Avila,
[25] J. A. Jaleel, S. Salim, and R. Aswin, “Computer aided detection and E. Valle, “Knowledge transfer for melanoma screening
of skin cancer,” in 2013 International Conference on Circuits, with deep learning,” in 2017 IEEE 14th International Sympo-
Power and Computing Technologies (ICCPCT), Nagercoil, sium on Biomedical Imaging (ISBI 2017), pp. 297–300, Mel-
India, March 2013. bourne, VIC, Australia, April 2017.
[26] A. Esteva, B. Kuprel, R. A. Novoa et al., “Dermatologist-level [41] C. N. Vasconcelos and B. N. Vasconcelos, “Experiments using
classification of skin cancer with deep neural networks,” deep learning for dermoscopy image analysis,” Pattern Recog-
Nature, vol. 542, no. 7639, pp. 115–118, 2017. nition Letters, 2017.
[27] K. Ramlakhan and Y. Shang, “A mobile automated skin lesion [42] R. B. Oliveira, A. S. Pereira, and J. M. R. S. Tavares, “Compu-
classification system,” in 2011 23rd IEEE International Confer- tational diagnosis of skin lesions from dermoscopic images
ence on Tools with Artificial Intelligence (ICTAI), Boca Raton, using combined features,” Neural Computing and Applica-
FL, USA, November 2011. tions, vol. 31, no. 10, pp. 6091–6111, 2019.
[28] A. N. Hoshyar, A. Al-Jumaily, and R. Sulaiman, “Review on [43] Y. Yuan, M. Chao, and Y. C. Lo, “Automatic skin lesion seg-
automatic early skin cancer detection,” in 2011 International mentation using deep fully convolutional networks with jac-
Conference on Computer Science and Service System (CSSS), card distance,” IEEE Transactions on Medical Imaging,
Nanjing, China, June 2011. vol. 36, no. 9, pp. 1876–1886, 2017.
[29] P. K. Goyal and M. K. Jain, “Computer-aided diagnosis of mel- [44] L. Bi, J. Kim, E. Ahn, and D. Feng, “Automatic skin lesion anal-
anoma skin cancer: a review,” in Advances in Data and Infor- ysis using large-scale dermoscopy images and deep residual
mation Sciences. Lecture Notes in Networks and Systems, vol 38, networks,” 2017, http://arxiv.org/abs/1703.04197.
M. Kolhe, M. Trivedi, S. Tiwari, and V. Singh, Eds., pp. 63–73, [45] M. A. Al-Masni, M. A. Al-antari, M. T. Choi, S. M. Han, and
Springer, Singapore, 2018. T. S. Kim, “Skin lesion segmentation in dermoscopy images
[30] A. Karargyris, O. Karargyris, and A. Pantelopoulos, “DER- via deep full resolution convolutional networks,” Computer
MA/care: an advanced image-processing mobile application Methods and Programs in Biomedicine, vol. 162, pp. 221–231,
for monitoring skin cancer,” in 2012 IEEE 24th International 2018.
Conference on Tools with Artificial Intelligence (ICTAI), Ath- [46] M. Jahanifar, N. Zamani Tajeddin, B. Mohammadzadeh Asl,
ens, Greece, November 2012. and A. Gooya, “Supervised saliency map driven segmenta-
[31] H. Xu, C. Lu, R. Berendt, N. Jha, and M. Mandal, “Automated tion of lesions in dermoscopic images,” IEEE Journal of Bio-
analysis and classification of melanocytic tumor on skin whole medical and Health Informatics, vol. 23, no. 2, pp. 509–518,
slide images,” Computerized Medical Imaging and Graphics, 2019.
vol. 66, pp. 124–134, 2018. [47] P. Tschandl, C. Sinz, and H. Kittler, “Domain-specific
[32] Y. Li and L. Shen, “Skin lesion analysis towards melanoma classification-pretrained fully convolutional network encoders
detection using deep learning network,” Sensors, vol. 18, for skin lesion segmentation,” Computers in Biology and Med-
no. 2, p. 556, 2018. icine, vol. 104, pp. 111–116, 2019.
14 BioMed Research International

[48] L. Bi, J. Kim, E. Ahn, A. Kumar, D. Feng, and M. Fulham,

“Step-wise integration of deep class-specific learning for der-
moscopic image segmentation,” Pattern Recognition, vol. 85,
pp. 78–89, 2019.
[49] S. Boltz, SRM method implementation, 2010, http://www
.mathworks.com/matlabcentral/fileexchange/authors/73145.
[50] R. Crandall, Level set implementation, https://github.com/
rcrandall/ChanVese. http://www.opencv.org..
[51] Q. Zhao, “JSEG method implementation,” 2001,
cs.joensuu.fi/Zhao/Software/JSEG.zip.
[52] M. D’Amico and I. Stanganelli, “Qualitative asymmetry
measure for melanoma detection,” in 2004 2nd IEEE Inter-
national Symposium on Biomedical Imaging: Nano to Macro
(IEEE Cat No. 04EX821), pp. 15–18, Arlington, VA, USA,
April 2004.
[53] T. Tanaka, R. Yamada, M. Tanaka, K. Shimizu, M. Tanaka,
and H. Oka, “A study on the image diagnosis of melanoma,”
in The 26th Annual International Conference of the IEEE Engi-
neering in Medicine and Biology Society, pp. 1597–1600, San
Francisco, CA, USA, September 2004.
[54] I. Maglogiannis and D. I. Kosmopoulos, “Computational
vision systems for the detection of malignant melanoma,”
Oncology Reports, vol. 15, pp. 1027–1032, 2006.
[55] J. S. Marques, C. Barata, and T. Mendonc, “On the role of
texture and color in the classification of dermoscopy
images,” in 2012 Annual International Conference of the
IEEE Engineering in Medicine and Biology Society,
pp. 4402–4405, San Diego, CA, USA, August-September
2012.
[56] J. I. Olszewska, “Semantic, automatic image annotation based
on multi-layered active contours and decision trees,” Interna-
tional Journal of Advanced Computer Science and Applications,
vol. 4, no. 8, pp. 201–208, 2013.
[57] E. Zagrouba and W. Barhoumi, “An accelerated system for
melanoma diagnosis based on subset feature selection,” Jour-
nal of Computing and Information Technology, vol. 13, no. 1,
pp. 69–82, 2005.
[58] M. Aljanabi, Y. Ozok, J. Rahebi, and A. Abdullah, “Skin lesion¨
segmentation method for dermoscopy images using artificial
bee colony algorithm,” Symmetry, vol. 10, no. 8, p. 347, 2018.