ArticlePDF Available

Authors:
Article

# hybSVM: Bacterial colony optimization algorithm based SVM for malignant melanoma detection

## Abstract and Figures

Melanoma is a malignant and aggressive type of skin cancer. This paper describes an effective method for detection of melanoma. A hybrid classification algorithm was developed by using the SVM algorithm and a heuristic optimization algorithm. In this algorithm, the SVM algorithm which uses a Gaussian Radial Basis Function (RBF) was enhanced by the Bacterial Colony algorithm (hybSVM). The model was tested with two different datasets namely ISIC and PH2 by using 10 cross fold validation. According to results AUC value of 98%, 97% and an operation time of 26.5, 11.9 sec obtained respectively from ISIC and PH2.
Content may be subject to copyright.
hybSVM: Bacterial colony optimization algorithm based SVM for
malignant melanoma detection
Sümeyya _
Ilkin
, Tug
˘rul Hakan Gençtürk, Fidan Kaya Gülag
˘ız, Hikmetcan Özcan, Mehmet Ali Altuncu,
Suhap Sßahin
Computer Engineering Department, Kocaeli University, Kocaeli 41001, Turkey
article info
Article history:
Revised 22 January 2021
Accepted 5 February 2021
Available online xxxx
Keywords:
Classiﬁcation
Skin lesion clustering
Machine learning
Malignom detection
Medical image processing
abstract
Melanoma is a malignant and aggressive type of skin cancer. This paper describes an effective method for
detection of melanoma. A hybrid classiﬁcation algorithm was developed by using the SVM algorithm and
a heuristic optimization algorithm. In this algorithm, the SVM algorithm which uses a Gaussian Radial
Basis Function (RBF) was enhanced by the Bacterial Colony algorithm (hybSVM). The model was tested
with two different datasets namely ISIC and PH2 by using 10 cross fold validation. According to results
AUC value of 98%, 97% and an operation time of 26.5, 11.9 sec obtained respectively from ISIC and PH2.
Ó2021 Karabuk University. Publishing services by Elsevier B.V. This is an open access article under the CC
1. Introduction
Melanoma is a type of skin cancer that starts in melanocyte
cells, which give its colour to the skin, and has a successful treat-
ment rate of 70% when diagnosed in an early stage. In our country,
there has been a 237% increase in melanoma cases in the last
30 years [1,2]. Melanoma is known to be the most aggressive type
of skin cancer [3]. It is mostly diagnosed by visual examination of
the skin rather than expensive and painful biopsies [4]. In this
method, which is also referred to as the ABCD rule (E feature is
added recently, so it is now called the ABCDE rule), Asymmetry,
Border irregularity, Color variation, Diameter, and lately, Evolving
of the spots or moles are evaluated on the dermoscopy images of
the skin [5]. It is a subjective evaluation as the method depends
on the expertise of the dermatologist or the oncologist [6], hence
the success of the diagnosis depends on the personal experience
of the physician [7]. For this reason, the early diagnosis of mela-
noma is still considered as a challenging and difﬁcult task for der-
matologists [8]. According to the studies, the sensitivity of
dermatologists for melanoma detection was 70% in Nevisense trial
and 78% in MelaFind trial [9]. Therefore, in the recent years, the use
of computer aided diagnosis (CAD) systems has become more pop-
ular, in order to increase the accuracy of diagnosis. While the ratio
of diagnosis in the early stages of melanoma cases is 60%, this ratio
is observed to range up to 90% with evaluations performed by dig-
ital dermatoscopy [2].
In the literature review, it is observed that many different
methods and techniques are used in the ﬁeld of malignant skin
lesion detection. Conoci et. al. [3] described an effective pipeline
for nevus analysis. The proposed pipeline was based on a combined
approach that uses ad-hoc customized image features and a feed-
forward Neural Network System (NNS). The described method was
also ported on an embedded system introducing a ‘‘Point of Care
(PoC)” hardware solution for high-speed nevus discrimination. This
study was promising with regards to its performance analysis.
Waheed et. al. [10] described a machine learning based method
for the detection of melanoma from dermoscopic images. In their
study, SVM classiﬁer was chosen to classify melanoma images
based on extracted color and texture features among all dermo-
scopic images. 13-D feature vector formed by including nine color
features and four texture features were used. According to the test
results, highest accuracy of 96% was achieved using both color and
texture features. Esteva et. al. [11] used a single CNN for classifying
skin cancer images from ISIC dataset. Used method was evaluated
according to obtained results with 21 clinical experts. Considering
the test results, CAD systems with deep neural networks achieved
the same performances as clinical experts. Yue et al. [12] proposed
a two-stage method for melanoma detection. They used deep CNN
method with more than 50 layers which are FCRN in the segmen-
tation phase and DRN in the classiﬁcation phase. ISBI dataset were
https://doi.org/10.1016/j.jestch.2021.02.002
2215-0986/Ó2021 Karabuk University. Publishing services by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
Corresponding author.
Ilkin).
Peer review under responsibility of Karabuk University.
Engineering Science and Technology, an International Journal xxx (xxxx) xxx
Contents lists available at ScienceDirect
Engineering Science and Technology,
an International Journal
journal homepage: www.elsevier.com/locate/jestch
Ilkin, Tug
˘rul Hakan Gençtürk, F. Kaya Gülag
˘ız et al., hybSVM: Bacterial colony optimization algorithm based SVM for
malignant melanoma detection, Engineering Science and Technology, an International Journal, https://doi.org/10.1016/j.jestch.2021.02.002
used for performance evaluation and gained 80.4% AUC. Carrera
and Ron-Dominguez [6] proposed a CAD system that detects mel-
anoma skin cancer using dermatoscopy images, SVM and DT clas-
siﬁer. The proposed system was tested by using 748 dermatoscopy
images and 28 features (two asymmetry, one border quality, 24
texture, and one blue-white veil) and according to the sensitivity
analyses, the success rate of the system was around 98% by SVM.
Mustafa and Kimura [13] proposed an automated system for
detecting melanoma. In their study, lesion images were segmented
by using the GrabCut algorithm for melanoma detection, and lesion
features such as the shape, color, and geometry were extracted.
These extracted features were categorized as ‘‘malignant” for
cancerous, or ‘‘benign” for non-cancerous mole by using SVM with
RBF kernel. The study was tested with 200 images and according to
the test results, only six features were found to be sufﬁcient for
melanoma detection. Pham et al. [14] proposed a deep CNN study
for melanoma classiﬁcation. They also contributed to the data aug-
mentation too. Each feature’s extraction process was done in dif-
ferent level layers of CNN. And dataset was artiﬁcially
augmented. The study was tested by using ISBI dataset and
obtained 89.20% AUC. Li and Shen [15] proposed two deep learning
methods for segmentation, feature extraction and classiﬁcation
processes of skin lesion analysis. In the paper, the FCRN were used
for segmentation process and LICU were used for calculate the dis-
tance map in the classiﬁcation process. The system was tested
using ISIC dataset and obtained 91.20% AUC value. Seeja and Sur-
esh [16] proposed a Deep CNN’s for accurate skin lesion segmenta-
tion using U-net algorithm. They combined Deconvolutional
network and Fully Connected Network. Color, texture and shape
features from the segmented images were extracted. Local Binary
Pattern (LBP) method were used for texture analysis. Edge his-
togram, Gabor and Histogram of Oriented Gradients (HOG) meth-
ods were also used for shape feature extraction. Then SVM, RF,
KNN and NBC classiﬁers were chosen for classiﬁcation. According
to the test results, The Dice co-efﬁciency value was 77.5% for image
segmentation and SVM classiﬁer’s accuracy value was 85.19%. Sur-
o
´wka and Ogorzalek [17] leaned on a wavelet-based feature
extraction process for melanoma detection with SVM classiﬁers
for dermoscopic skin images. 53 wavelet bases were examined in
order to determine which gave the best results for SVM. The study
was tested with two different datasets that have different resolu-
tions and sizes. Murugan et al. [18] proposed a classiﬁer for mela-
noma detection. Watershed method were used for segmentation
step and later these results used in the feature extraction process.
They tested the proposed approach with 3 different classiﬁcation
methods which are SVM, KNN and RF. According to test results
SVM gave 89.43% accuracy value. Alsaeed [19] proposed a SVM
based CAD system for skin cancer detection. In the paper, SVM
was used with texture, shape and color features. Totally 20 features
were used. System was tested with PH2 and UMCG datasets. The
two HCI experts also tested the system. Hosny et al. [20] proposed
a deep learning-based skin lesion classiﬁcation system. In the
paper, Alex-net have been used for ﬁne-tuning the weights. The
alteration of the classiﬁcation layer of Alex-net were done by using
the softmax. DCNN weights were proposed and tested their study
with MED-NODE, Derm and ISIC datasets. Nasiri et. al. [21] devel-
oped a case-based reasoning system for early detection of mela-
noma. Their approach depended on deep learning technique for
classifying the skin lesions. This study is actually a further study
of their DePict Melanoma Class. In their previous study, it used
region growing methods based on SVM and k-NN to retrieve the
textual elements and classify melanoma images. In this study a
deep learning technique which is a 19-layer model CNN was used
for classifying the skin lesions. The model has eleven convolu-
tional, ﬁve max-pooling and three fully connected layers. The pro-
posed algorithm was tested using the ISIC Melanoma Project and
obtained 75% success rate. Astorino et al. [22] proposed a MIL tech-
nique for melanoma detection. Proposed system tested with PH2
dataset with 10 cross fold validation and obtained 90.63% accuracy
rate. Arora et al. [23] focused on feature extraction process of skin
cancer detection in their study. Bag of feature method were used
for feature extraction and quadratic SVM for classiﬁcation. The
tests of the study were performed using the PH2 dataset and an
accuracy of 85.70% was obtained. When the studies are analyzed,
it can be seen that there are numerous methods used to detect skin
lesions with melanoma. In the studies that use classiﬁcation, SVM
classiﬁer, which is frequently used in studies with a melanoma
data set, is observed to achieve successful results in medical image
analysis [24]. Also new approaches such as deep learning has
recently used for various medical image classiﬁcation problems
as well. Połap et al. [25] proposed a smart home system for detect-
ing the skin conditions especially the nevus by using convolutional
neural network. Zunair and Hamza [26] proposed two-stage frame-
work for skin lesion classiﬁcation which using deep convolutional
neural network in the classiﬁcation step. Additionally, bio-inspired
heuristic algorithms such as ﬁreﬂy, artiﬁcial bee colony, artiﬁcial
ant colony, moth ﬂame and particle swarm algorithms has started
to use in medical image analysis, like Wo
´zniak and Połap [27],
Akkar and Salman [28].
It was observed that system performances were usually
assessed at the segmentation or classiﬁcation steps through vari-
ous methods. However, it can be seen that different methods were
not employed at either step and although SVM was mostly used in
the classiﬁcation step, there were not many studies conducted to
improve SVM operation times. Therefore, in our study, we devel-
oped a hybrid system which yields the most successful classiﬁca-
tion results in melanoma skin lesions. In the segmentation step
of our study FSLIC [29] method used. The performance of the seg-
mentation step was calculated by the Dice, Jac, Accuracy, F-
measure, Mcc metrics. As a result of the calculation, it was con-
cluded that FSLIC with a k-center value of 4 was the best lesion
segmentation method. Our data set was tagged with FSLIC before
proceeding to the classiﬁcation step. RC, NBC, DT, SVM,
[6,10,13,16] machine learning techniques which were observed
to be used frequently in studies were used in the classiﬁcation step
of our study. Finally, to increase the efﬁciency of the SVM algo-
rithm, a hybrid machine learning algorithm hybSVM was devel-
oped by using the bacterial colony algorithm, which is one of the
heuristic optimization algorithms. Accuracy, MSE, Precision, Recall,
F-measure metrics were calculated to measure the effectiveness of
the study. ROC curve and AUC of the system were calculated with
these values. When AUC values were analyzed, it was concluded
that the best malign skin lesion detection procedure was obtained
through FSLIC segmentation and hybSVM.
In Section 2, the architecture of the study and the methods used
in the study are explained in detail. In Section 3, hybSVM, which
was developed within the scope of the study is described in detail.
In Section 4, the evaluation metrics used in the study are described.
In Section 5, Results of the tests performed within the scope of the
study are presented and the results are interpreted. In Section 6,a
general evaluation within the scope of this study and improve-
ments are presented.
2. System architecture and methods used
The system architecture shown in Fig. 1 is represented below in
detail.
In the ﬁgure, ﬁrst a training model is established with the ROI
(Region of Interest) images in our data set through SVM. Our train-
ing model is recorded for later use. FSLIC clustering method is uti-
lized in the segmentation phase of the study. Our dataset is
Sümeyya
_
Ilkin, Tug
˘rul Hakan Gençtürk, F. Kaya Gülag
˘ız et al. Engineering Science and Technology, an International Journal xxx (xxxx) xxx
2
clustered by using FSLIC method. The clusters obtained through the
clustering method are labelled by the trained model. By combining
the clusters with the same label on the images, the number of clus-
ters in each image is reduced to 2, as 0 (areas without melanoma)
and 1 (areas with melanoma). ROI images are generated based on
the label values of the images, and our resulting images are
recorded by calculating the metric analyses based on these images.
Finally, the hybSVM algorithm developed through the bacterial
colony algorithm is executed. These steps are described in detail
below.
2.1. Segmentation
Segmentation used in the study is commonly used to eliminate
background noises (normal skin texture, ﬁlming ﬂares, skin hair
etc.) on the images. More successful classiﬁcation results are
achieved by using lesion ROI images obtained by segmentation.
Since these ROIs directly affect system performance, the algo-
rithms and output selected at this step are of great signiﬁcance.
Therefore, FSLIC method which is frequently used in medical image
processing studies was used to achieve the most successful result
in the segmentation stage of our study.
Clusters are formed with the SLIC method by clustering pixel
groups based on the similarity and proximity of the color values
of images. These color-based clusters are referred to as superpixels.
CIELAB color space is used for this process. The desired number of
clusters is speciﬁed with the k parameter used in the algorithm and
superpixel clusters are established equally according to the value
of this parameter [30,31].
The FSLIC method, which was developed by using the SLIC
method and gave faster results, was used in our study. The devel-
oper of this method succeeded to execute the SLIC method faster
and more effectively by using a CPU. He provided the method he
developed as an open source code in the github code sharing net-
repository [32].
In order to determine the effectiveness of the study in the
method, we produced results by specifying 4 different k center val-
ues. When previous studies are examined, there are no descrip-
tions or rationale about the selection of the k-center value [33].
Hence, the users usually choose the k center value based on their
experience in datasets [34,35]. Within the scope of this study, it
was deemed appropriate to select the k-center values used in the
FSLIC clustering algorithm as 4, 8, 16 and 32, since the background
values were not too signiﬁcant to affect the results or generate
noise in the macroscopic images. Images obtained as a result of
the segmentation operations performed with four different k cen-
ter values are recorded. These images are labelled by our trained
model. Then the images with the same label are consolidated.
2.2. Label merging
For the classiﬁcation step, there must be 2 classes. For this aim,
each cluster obtained as a result of segmentation is labelled as 1 or
0 according to the training model. Label clustering is performed to
form 2 clusters out of the clusters obtained after the segmentation
step. From these image sets, which are classiﬁed into 2 clusters by
labelling each image as 0 and 1, clusters with the same label are
consolidated and converted to binary. As a result of these opera-
tions, the number of clusters on each image is reduced to 2. The
pseudocode for this process is represented in Algorithm 1.
Algorithm 1. The pseudocode of the cluster combining
/* Assignment */
Initialize cluster centers knum = [4,8,16,32] by using each
value in a different clustering process
Set assignment = slic.iterate(image) for obtain to the cluster
map.
Set cluster_rgb = [[] for i in range(knum)] for dividing the
cluster’s rgb values into clusters.
Set centers = [] for obtain cluster’s center color values.
for each cluster cluster_rgb do
Compute cluster mean, std, var, skew, kurtosis values.
if svmclassiﬁer.predict == 0 then
centers[i]=(0,0,0)
else
centers[i]=(255,255,255)
endif
end for
/* Update */
Set new clusters.
Set image as a binary image.
After these operations, ROI images are generated by using the label
values of the images. Fig. 2 shows sample ROI images.
2.3. Feature extraction
At this stage, the features of the original images in the dataset
and their ROI images produced by the specialist are extracted to
be used in classiﬁcation. In our study, 9 feature values for the
lesion areas are extracted in accordance with the ABCDE melanoma
detection rule [5].
1) Colour
The Mean (m), Standard Deviation (
r
), Variance (
r
2
N
), Skew-
ness (b
S), Kurtosis (b
K) features of the colour channel values
of the images are calculated and shown respectively in
Eqs. (1)–(5).
Fig. 1. The architecture of the system developed.
Fig. 2. Examples of referenced images a) original image b) ROI image.
Sümeyya
_
Ilkin, Tug
˘rul Hakan Gençtürk, F. Kaya Gülag
˘ız et al. Engineering Science and Technology, an International Journal xxx (xxxx) xxx
3
EX½¼
l
¼XxPxðÞ
l
¼1
nX
n
i¼1
xiðÞ¼X
n
i¼1
xN
x
n
PxðÞ¼lim
n!1
N
X
n
so lim
n!1
l
¼PxPxðÞ¼EX½ therefore mean for
image features represented by:
l
¼1
NX
N
i¼1
S
i
x;yðÞ ð1Þ
In Eq. (1), (x, y) 2W represents the rows and columns which is
the average of summation all pixels of the image and N
x
value rep-
resents the repeating count.
r
¼ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
EX
l
ðÞ
2
hi
rtherefore standard deviation for image fea-
tures represented by:
r
¼ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
1
NX
N
i¼1
S
i
x;yðÞ
l
ðÞ
2
v
u
u
tð2Þ
In Eq. (2), (x, y) 2W represents rows and columns which shows
the dispersion occurs from the average pixel value.
r
¼ﬃﬃﬃﬃﬃﬃ
r
2
N
q
r
2
N
¼EX
2
hi
l
2
r
2
N
¼Px
2
PxðÞ
P2
l
xP xðÞþ
P
l
2
PxðÞ¼EX
2
hi
l
2
therefore
variance for image features represented by:
r
2
N
¼1
NX
N
i¼1
S
i
x;yðÞ
l
ðÞ
2
ð3Þ
In Eq. (3) gives us the information about how far the pixel
points are spread from the average value.
b
S¼1
NX
N
i¼1
Six;yðÞ
1
NPN
i¼1Six;yðÞ
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
1
NP
N
i¼1
Six;yðÞ
1
NP
N
i¼1
Six;yðÞ

2
sÞ
3
¼1
NX
N
i¼1
Six;yðÞ
l
r

3
0
B
B
B
B
@ð4Þ
b
K¼1
NX
N
i¼1
Six;yðÞ
1
NPN
i¼1Six;yðÞ
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
1
NP
N
i¼1
Six;yðÞ
1
NP
N
i¼1
Six;yðÞ

2
sÞ
4
¼1
NX
N
i¼1
Six;yðÞ
l
r

4
0
B
B
B
B
@ð5Þ
In the Equations above, S(x, y) represents the segmented area,
S
i
(x, y) takes values in the set S. W represents the segmented
region and N represents the total number of pixels of the images
[36].
2) Shape
One of the key features used for the detection of melanocyte
lesions is their irregular shape according to the ABCDE rule
[37]. If the shape of the lesion is regular, the ratio is 1, other-
wise the ratio nearer to 0 [18]. To obtain the shape informa-
tion of the lesions, the Asymmetry, Border Irregularity (I),
Compactness (C) and Red Relative (RR) values of the images
are calculated.
a) Asymmetry
Asymmetry feature of the images is computed as shown
in Eq. (6).
Assy ¼
D
AK
AL 100 ð6Þ
In Eq. (6),
D
AK represents the area of between two lesion halves
and AL represents lesion area.
b) Border Irregularity
Border irregularity feature of the images is computed as
shown in Eq. (7).
I¼xy
2
p
x
2
þy
2
ðÞ
P
2
D
Að7Þ
In Eq. (7), x and y represent distance of the major and minor
axis points of the lesion and P represents the lesion perimeter
value, whereas
D
A represents the corresponding area.
c) Compactness
Compactness feature of the images is computed as shown in
Eq. (8).
C¼4
p
P
2
D
Að8Þ
In Eq. (8), P represents the lesion perimeter value, whereas
D
A
represents the corresponding area.
d) Red Re lative
Red relative feature of the images is computed as shown in
Eq. (9). Since segmentation is performed in our study, skin
value is eliminated when the relative feature is extracted
and only the relative value of the red color is used.
RR ¼r
L
r
L
þg
L
þb
L
ð9Þ
In Eq. (9),r
L
,g
L
and b
L
represent the mean red, green and blue
appearing within the lesion [37].
2.4. Establishment of training data
If we assume that the values obtained as the result of the calcu-
lations given above is matrix M(x, y); x is the image ID number,
and y is the feature value of this image. To be able to use the image
features in the classiﬁcation operations, the feature matrix is con-
verted to a vector. This is done by taking the feature values of all
images one by one and generating a single-dimensional feature
vector. Our feature vector obtained in this manner is labelled by
using the ROI. A training data is generated with these labelled vec-
tors. Our training data are randomly divided into two to be used in
classiﬁcations; 80% as training educational data and 20% as test
data.
2.5. Classiﬁcation
The generated training data are separately trained by using
SVM, RC, NBC and DT classiﬁers.
2.5.1. Support Vector Machine (SVM)
SVM is a supervised learning method which is often used for
regression and classiﬁcation problems. It was developed by Vapnik
as a machine learning method. Based on Lagrange multipliers, SVM
is a method that aims to ﬁnd the optimum separator to separate
two different classes [31]. SVM provides a uniﬁed framework in
which different data can be classiﬁed through a suitable kernel
selection. We can consider this as one of the key advantages of
SVM [16]. We use a linear kernel SVM to classify our dataset in
Sümeyya
_
Ilkin, Tug
˘rul Hakan Gençtürk, F. Kaya Gülag
˘ız et al. Engineering Science and Technology, an International Journal xxx (xxxx) xxx
4
the ﬁrst step. The most important reason for choosing the linear
kernel is its speed and accuracy in binary classiﬁcation problems.
In the comparison stage, two different SVM software toolkits
which are Scikit-learn and ThunderSVM is used for demonstrate
the effectiveness of the proposed method. Scikit-learn is a Python
integration module which is developed for implemented in high-
level languages. Scikit-learn presents a wide range of variety for
implementing supervised and unsupervised machine learning
methods [38]. ThunderSVM is an open source SVM software
toolkit. It is built and used for getting higher performance for pro-
ducing identical SVMs. This software toolkit exploits the GPUs and
multi-core CPUs while executes the SVMs [39]. ThunderSVM fasts
the SVM computation time by parallelizing the kernel computation
steps of SVM in the CPU level [40].
2.5.2. Random Forest Classiﬁer (RC)
Random Forest is a popular ﬂexible machine learning algorithm,
which is established with a number of decision trees [18]. Each
decision tree was created at the time of training and grown using
a randomization form and outputting the class of the individual
trees [41].
2.5.3. Naïve Bayes Classiﬁcation (NBC)
NB classiﬁer uses the Bayes theorem for classiﬁcation problems.
NB classiﬁers assume that all variables are strong or independent
between attributes of data points in the classes. These independent
variables are known as naïve because they are not suitable to the
real world [16,42]. NB classiﬁers are commonly used for data min-
ing problems [43].
2.5.4. Decision Tree Classiﬁer (DT)
Decision Tree is one of the supervised machine learning algo-
rithms. It is a simple classiﬁcation method which splits data
according to their attributes [16]. It is called a tree because of its
branching feature. DT classiﬁcation has 3 tree terms which are
nodes, branch and leaf nodes. The differences of these 3 terms
are; nodes indicate the test results of the feature which are used
for the problem, branch denotes the outcome of the test and leaf
nodes indicate the class label [44].
Classiﬁcation results are obtained by training our dataset
labelled by each machine learning algorithm described above.
These results are tested with the test data separated from our data-
set in a ratio of 20%. The efﬁciency of the study is measured by
metric calculations performed with the values obtained as a result
of these operations. In the ﬁnal step, we start with the enhance-
ment process on SVM to increase the effectiveness of the SVM
algorithm.
3. Proposed enhanced SVM algorithm (hybSVM)
Two parameters of the SVM algorithm that signiﬁcantly affect
the effectiveness of the SVM are gamma and C. The gamma param-
eter describes the distance effect of a single training sample. Low
values represent that the impact of the training example is far, high
values represent that the impact of the training example is close.
The C parameter works like the regularization parameter in the
SVM algorithm. It changes the classiﬁcation of training instances
based on maximizing the margin of the decision function. The lar-
ger of the C value, margin is considered as the smaller if the deci-
sion function can correctly classify all training points. The smaller
of the C value, margin is accepted as the larger to affect the training
accuracy. Hence, while the C value is getting smaller, it promotes
the simpler decision function [45]. Therefore, the choice of gamma
and C parameters has a great importance. Various studies are car-
ried out to ﬁnd the best combinations that maximizes the perfor-
mance in parameter choice. These studies are considered as an
optimization problem and search algorithms are frequently used
in them [46]. In our study, we used a heuristic optimization algo-
rithm, which is a bacterial colony optimization algorithm to obtain
the best combination of the gamma and C parameters of SVM.
Bacterial Colony Optimization Algorithm is a computation tech-
nique that is developed to solve complex engineering problems
inspired by the foraging behavior of E. coli bacteria. It is basically
made up of 3 events: chemotaxis, reproductive and elimination-
dispersal [47]. The basic steps of the bacterial colony algorithm
are shown in Algorithm 2.
Algorithm 2. Bacterial Colony Algorithm Basic Steps
Step 1: Generate n bacteria at random.
Step 2: Evaluate bacteria according to the evaluation
function.
Step 3: Set up three cycles for optimization:
Step
3.1:
Internal loop: Chemotactic event
Step
3.2:
Middle cycle: Reproductive phenomenon
Step
3.3:
Outer loop: Elimination-disintegration event
Step 4: Show the best global results at the end of the
iterative process.
The ﬂowchart of the developed hybSVM algorithm is shown in
Fig. 3. The hybSVM algorithm developed within the scope of the
study executes in 2 phases.
In the ﬁrst phase; the gamma value used in the RBF SVM algo-
rithm is modiﬁed according to the optimum acc value by drawing a
route with the bacterial colony algorithm. The pseudocode of the
ﬁrst phase of the hybSVM algorithm is presented in Algorithm 3.
Algorithm 3. The pseudocode of the hybSVM
/* Assignment */
Set calculateDistance(g):
Set SVM = (kernel=’rbf’, random_state = 0, gamma = g, C = 32)
return accuracy_metricvalue(y_test,y_pred)
Set kemotaxis(bacteria): for kemotaxis procedure
Set bacteria = bacteria[0]
Calculate bacteria distance
Set elimination(bacteria): for elimination procedure
Set bacteria = bacteria[0]
Calculate bacteria distance
/* Update */
numBacteria = 4
maxGen = 20
for i in range(numBacteria):
gamma = 2**(numBacteriaHalf-i)
bacteria.append((gamma, calculateDistance(gamma)))
/*Execute*/
Choose low cost bacteria for migration
Kemotaxis procedure
Re execute low cost bacteria
Bacteria reproduction
/*Update*/
In the algorithm, ﬁrst, a randomly chosen C value is taken as a con-
stant and the initial gamma value is generated. In the bacterial col-
ony algorithm, bacterial movements are produced by multiplying
Sümeyya
_
Ilkin, Tug
˘rul Hakan Gençtürk, F. Kaya Gülag
˘ız et al. Engineering Science and Technology, an International Journal xxx (xxxx) xxx
5
with a value between 0.01 and 2. The improvement process is per-
formed in 20 iterations with 4 bacteria which are determined from
the test results given in Section 5. Bacteria produced for bacterial
elimination are jumped back or forth in half. This means that the
produced bacteria can pass to the next step in big jumps with a
50% probability. The costs of the produced bacteria are calculated,
and chemotaxis process is performed to strengthen the bacteria.
Reproduction is performed to reprocess the bacteria with poor
costs. Then, whether the elimination process will be performed is
decided with a probability of 0.01%. Hence, the bacteria which yield
the best cost by the end of 20 iterations are selected and the gamma
value is modiﬁed according to the accuracy value. After the best
gamma value is obtained in the ﬁrst phase, we proceed to the sec-
ond phase.
In the second phase, the best gamma value is taken as a con-
stant and the bacterial colony algorithm is executed again to obtain
the best C value. After these two phases, the best gamma and C val-
ues are obtained. This way, SVM is executed with the gamma and C
values that yield the best costs with the hybSVM algorithm.
4. Evaluation metrics
To validate the performance and effectiveness of the study,
metric calculations were performed on the results of the methods
used in the segmentation and classiﬁcation steps of the study.
Fig. 3. The ﬂowchart of the developed hybSVM algorithm.
Fig. 4. Sample images used in the testing phase from ISIC dataset a) original image
b) ROI from dataset.
Fig. 5. Sample images used in the testing phase from PH2 dataset a) original image
b) ROI from dataset.
Table 1
FSLIC Segmentation results.
Metric K = 4 K = 8 K = 16 K = 32
Accuracy 0.960 0.939 0.936 0.933
F-measure 0.929 0.890 0.884 0.879
Dice 0.929 0.890 0.884 0.879
Jac 0.864 0.795 0.778 0.766
Mcc 0.902 0.848 0.841 0.833
Table 2
Prediction Results of RF, NBC, DT, SVM Classiﬁers.
Metric SVM DT NBC RF
Accuracy 0.97 0.91 0.92 0.95
MSE 0.15 0.30 0.26 0.20
Precision 0.97 0.90 0.91 0.96
Recall 0.97 0.91 0.95 0.95
F-measure 0.97 0.91 0.93 0.95
Sümeyya
_
Ilkin, Tug
˘rul Hakan Gençtürk, F. Kaya Gülag
˘ız et al. Engineering Science and Technology, an International Journal xxx (xxxx) xxx
6
In metric operations, True Positives (TP) values indicate cor-
rectly segmented lesion pixels, True Negatives (TN) values indicate
correctly segmented background pixels, False Positives (FP) values
indicate background pixels classiﬁed as lesions and False Negatives
(FN) values indicate lesion pixels classiﬁed as background. The
Sensitivity metric measures the accuracy of the lesion segmenta-
tion. It is calculated as shown in Eq. (10).
Sen ¼TP
TP þFN ð10Þ
The accuracy of the segmentation of the background is mea-
sured with the speciﬁcity metric. It is calculated as shown in Eq.
(11).
Spec ¼TN
TN þFP ð11Þ
The accuracy metric is calculated as shown in Eq. (12). The
higher this value, the more successful is the study. Accuracy rate
demonstrates how accurate the study performs the classiﬁcation.
Precision, recall and f-measure metric values are calculated
through Eqs. (13)–(15), respectively.
Acc ¼TP þTN
TP þTN þFP þFN ð12Þ
Precision ¼TP
TP þFP ð13Þ
recall ¼TP
TP þFN ð14Þ
fmeasure ¼2:Precision:recallðÞ
Precision þrecall ð15Þ
The Jaccard Index (Jac) is a statistic metric. It measures similar-
ity between ROI and segmentation results by methods we use, and
it is calculated through Eq. (16). The higher the Jac value, the closer
the ROI value obtained through segmentation is to the ROI value
provided by the specialists.
jac ¼TP
TP þFP þFN ð16Þ
Such as the jac metric, dice similarity coefﬁcient (Dice) mea-
sures similarity [48]. The difference is that it gives more sensitive
and less weight outlier results than jac. It is computed as shown
in Equation (17).
Dice ¼2:TP
2:TP þFP þFN ð17Þ
The Matthew Correlation Coefﬁcient (Mcc) metric is calculated
through Eq. (18) [49].
Mcc ¼TP:TN FP:FN
ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
TP þFPðÞ:TN þFNðÞ:TP þFNðÞ:TN þFPðÞ
pð18Þ
5. Results and discussion
In the test phase of the study, two different open access datasets
are used which are the International Skin Imaging Collaboration
(ISIC) and PH2.
The ISIC is an open access dataset generated within the scope of
ISIC Melanoma Project provides a large data set made up of der-
moscopy images for research and educational purposes [50].
Images from this dataset consist of lesions of nevus, seborrhoeic
keratosis and malignant melanoma types in various diameters
and sizes [51]. PH2 dataset were created real case dermoscopy
images of the dermatology service of Hospital Pedro Hispano and
Fig. 6. Confusion matrix graphics of a) RF, b) NBC, c) DT, d) SVM.
Sümeyya
_
Ilkin, Tug
˘rul Hakan Gençtürk, F. Kaya Gülag
˘ız et al. Engineering Science and Technology, an International Journal xxx (xxxx) xxx
7
the research group of the University of Porto, Técnico Lisboa in
Matosinhos [52]. PH2 is an open access dataset generated within
the scope of Automatic computer-based Diagnostic system for Der-
moscopy Images Project (ADDI) which consists of 80 common nevi,
80 atypical nevi and 40 melanomas images [10]. The study was
tested by 10 cross fold validation with 1000 images from the ISIC
and 200 images from the PH2 datasets, separately. Those images
are of different shape and dimensions from 540 722 to
4499 6748 pixels [52] and Figs. 4 and 5 shows some sample
images respectively from datasets.
At this stage of our study; the ROI images obtained through the
FSLIC (k = 4, 8, 16 and 32 values) methods and the ROI values from
the data set which are shown in Fig. 4.b are compared, and metric
calculations are performed. The values obtained as a result of the
metric calculations described above is shown in Table 1.
When the results in Table 1 is analyzed, it can be concluded that
successful results are obtained as the number of k centers
decreases in the FSLIC method. Considering the test results, it can
be concluded that the selected k center values must be at the min-
imum value for the study in order for the FSLIC method to yield
successful results. It is concluded that the segmentation process
performed by a k-center value of 4 yields more successful results
with FSLIC method.
To assess the effectiveness of the study, accuracy, MSE, preci-
sion, recall, f-measure metrics are calculated on the data obtained
from 4 different classiﬁcation algorithms. The results of these cal-
culations are presented in Table 2.
When Table 2 is analyzed, it is concluded that the most success-
ful classiﬁcation results are obtained by the SVM classiﬁcation
algorithm. Fig. 6 presents confusion matrix graphics of classiﬁer’s
test results.
ROC curve values are calculated by the metric values obtained
to measure the effectiveness of the study. ROC curve is one of
the most important evaluation metrics for assessing the perfor-
mance of any classiﬁcation model [53].
The area under the ROC curve is calculated as the AUC (Area
Under Curve) value. The closer the AUC value is to 1, the more suc-
cessful is the classiﬁcation. Fig. 7 shows the ROC curve graphs of
the classiﬁcation methods. AUC values are shown in Table 3.
When Table 2 and 3 are analyzed together, it is concluded that
the most successful classiﬁcation result is obtained through the
SVM algorithm. As a result of these analysis operations, it is con-
cluded that the SVM algorithm yields considerably successful
results in lesion classiﬁcation processes. Therefore, SVM classiﬁca-
tion algorithm was selected to be used in the hybrid classiﬁcation
algorithm established within the scope of the study.
The hybSVM algorithm was tested with different bacteria and
iteration parameter values of Bacterial Colony algorithm to obtain
the best results. The test results are given in Table 4 and the com-
parison of the metrics graphs is shown in Fig. 8.
In Fig. 8.a., it can be seen that Precision, F-Measure, Accuracy
metrics results piked when the bacteria number is 4. After this
parameter value, all of three metrics results did not changed. Also,
Fig. 8.b. shows the performance of the algorithm according to iter-
ation numbers. When Fig. 8.b. is analyzed, it can be seen that for
the iteration number, 20 is the peak value according to all metrics.
There was no change in the performance of the algorithm in itera-
Fig. 7. ROC curve graphics of a) RF, b) NBC, c) DT, d) SVM.
Table 3
AUC Results of RF, NBC, DT, SVM Classiﬁers which are obtained from ROC curve
values.
SVM DT NYB RF
AUC Value 0.99 0.94 0.98 0.99
Sümeyya
_
Ilkin, Tug
˘rul Hakan Gençtürk, F. Kaya Gülag
˘ız et al. Engineering Science and Technology, an International Journal xxx (xxxx) xxx
8
tions above 20. These results show that saturation points of
hybSVM algorithm are 4 for bacteria number and 20 for iteration
number. For this reason, the ideal bacteria number for hybSVM
was determined as 4 and the number of iterations as 20.
The results of comparison to measure the effectiveness of the
hybSVM algorithm developed within the scope of the study with
the SVMs and ThunderSVM algorithms are shown in Table 5. The
confusion matrix and ROC curve graphs are shown in Fig. 9.
When Table 5 is examined, it can be seen that hybSVM per-
formed better than SVM and ThunderSVM in several metrics
shown in the table. In Table 5 and 6, Time column demonstrates
the execution time for all images included in the datasets. Scikit-
learn and ThunderSVM libraries were calculated the execution
time by summing up training time, model creation time and test-
ing phase time for all images. In hybSVM, unlike the other libraries,
Bacterial Colony Algorithm execution time was also included to
total execution time. In Table 5, the time criterion demonstrates
that the developed hybSVM algorithm yields best results with a
time period of less than ~3 times with SVM (Scikit-learn) and
~10 times with SVM (ThunderSVM). As shown in the confusion
matrix in Fig. 9.a. and e., the difference in the number of incorrect
labellings between the two algorithms is only 1. Considering the
small difference in the number of errors based on MSE rates and
confusion matrices, and the big difference in operation times
which is over threefold and tenfold, it is concluded that the
hybSVM algorithm developed within the scope of the study is
rather successful in the detection of melanoma. Performance tests
of the proposed algorithm were carried out on two different data
sets separately and shown in Table 6.
Table 6 demonstrates the overall results for common nevus,
atypical nevus, seborrhoeic keratosis and malignant melanoma
types. The results of the State-of-art comparison performed by 10
cross fold validation for ISIC and PH2 datasets are presented in
Table 7. In addition, the results of the State-of-art comparison to
measure the effectiveness of our hybSVM algorithm for ISIC and
PH2 datasets and are shown in Table 8.
In [18], they were focused on the segmentation and feature
extraction process of melanoma detection. Wide range of feature
selection methods which included shape, ABCDE and GLCM rules
were also used. When the feature extraction rules were tested, it
was observed that ABCDE rules work best with SVM. In [19], the
image sizes were ﬁxed to 575*765 px and the black corners were
cropped. Even that operations did not sufﬁce to enough. In [22]
multiple instance learning technique were used for classifying
the melanoma images and were obtained 92.50% accuracy rate.
In [23] put emphasis on feature extraction phase of the mela-
nom/skin cancer detection systems. In this study BoF via SURF fea-
ture extraction method with quadratic SVM was used [17] is also
another study which leaned on feature extraction process. They
were used 53 wavelet bases for detect the best match with
C-SVM. Also, SVM kernels were optimized by using Bayesian
search method. Yet 94.00% AUC rate was obtained. There were
Table 4
Performance test results according to iteration and bacteria number parameters of hybSVM algorithm.
Bacteria (#) Iteration (#) Time (sec.) Recall (%) Precision (%) F-Measure (%) Accuracy (%) Gamma C
2 20 24.1 98.40 96.10 97.24 97.20 1.23E + 08 10432
3 20 34.3 98.40 96.48 97.43 97.40 3.33E + 10 14270
4 20 38.5 97.94 96.95 97.44 97.50 4.79E + 04 14850
5 20 96 97.94 96.95 97.44 97.50 3.03E + 10 14911
6 20 98 97.94 96.95 97.44 97.50 2.58E + 04 13007
7 20 112 97.94 96.95 97.44 97.50 2.03E + 05 11027
8 20 135 97.94 96.95 97.44 97.50 2.00E + 10 14245
4 10 26.1 99.20 95.40 97.26 97.20 4.84E + 10 11024
4 15 35.2 98.80 96.12 97.44 97.40 5.01E + 10 14201
4 20 38.5 97.94 96.95 97.44 97.50 4.79E + 04 14850
4 25 72 97.94 96.95 97.44 97.50 5.03E + 05 12630
4 30 84 97.94 96.95 97.44 97.50 4.97E + 08 12493
Fig. 8. Precision, F-Measure, Accuracy metrics comparison graphics of the hybSVM
algorithm for a) iteration number parameter 20, b) bacteria number parameter 4.
Table 5
Comparison results of SVMs and hybSVM.
Method Time (sec.) MSE (%) Recall (%) Precision (%) F-Measure (%) Accuracy (%) AUC (%)
SVM (Scikit-learn) 67.00 15.80 97.43 97.43 97.43 97.50 99.00
SVM (ThunderSVM) 243.00 24.49 94.00 94.00 94.00 94.00 98.00
hybSVM 26.50 15.81 97.94 96.95 97.44 97.56 98.00
Sümeyya
_
Ilkin, Tug
˘rul Hakan Gençtürk, F. Kaya Gülag
˘ız et al. Engineering Science and Technology, an International Journal xxx (xxxx) xxx
9
two different datasets in the paper and we chose test results that
obtained on PH2 dataset to compare with our proposed algorithm.
When [18], [17], [19], [22] and [23] studies are examined, it has
been focused on subjects such as feature extraction, image
enhancement, application of different classiﬁcation techniques
and kernel optimization in order to increase the success for mela-
noma detection. When the methods mentioned above is compared
with the proposed hybSVM algorithm, it is seen that the AUC and
accuracy values obtained in the studies are lower than the pro-
posed algorithm.
In [11] a single CNN used for classifying the skin cancer lesions.
In the study Google Inception v3 architecture were used. When we
examine their test results their accuracy and AUC values underper-
form our test results even without discussion of execution times of
the studies. In [12] a two-phase CNN were used for melanoma
detection. In the study’s segmentation phase and classiﬁcation
phase was employed FCRN and DRN which are time consuming
deep learning methods. But in the training part were done by using
SVM. Moreover, the image sizes were ﬁtted to one small sizes. The
study was actually carried out using the ISBI dataset, but the
source of this dataset is the ISIC dataset. Hence, we decide to com-
pare this study with our proposed one. We used the SVM results
included in the paper to compare the results with our study. In
[14] were used a deep CNN. In the study, melanoma image feature
extraction was done in a layer-based approach. This adopted
approach were not give a satisfying result even without consider
to execution time. Again, we picked the SVM results obtained in
this study to compare with our proposed algorithm. In [15] too,
were focused on the feature extraction process of melanoma
images by using straight-forward CNN. In [20], authors augmented
the data for increase the system performances and changed the
Alex-net’s classiﬁcation layer with the softmax layer. And even,
the ﬁne-tuned weights and improved back-propagation still could
not sufﬁcient enough. In [52] presented a full resolution convolu-
tional network for skin lesion segmentation purpose. They tested
their study with two well-known datasets which are ISBI 2017
challenge and PH2 datasets. In [54], YOLO version 3 with GrabCut
algorithm is used for skin lesion segmentation. Also, in [55] a deep
learning algorithm is used which is artiﬁcial neural network that
trained with differential evolution algorithm for the melanoma
detection. In [56], authors were designed a hardware for mela-
noma detection and they used a combination of artiﬁcial neural
network and multilayer perceptron. In both studies tests carried
out on PH2 dataset too.
Even tough deep learning algorithms were found promising due
to their automatic feature extraction ability, they have found as a
very complex and time-consuming methods according to the stud-
ies [57–59]. In the proposed algorithm, we enhanced the classic
RBF kernel of SVM with a heuristic optimization algorithm and
the results achieved with 98% AUC rate for ISIC and 97% AUC rate
for PH2 datasets. Test results is shown that hybSVM is a beneﬁcial
approach both in terms of performance and time consumption.
Based on the obtained high accuracy rates, it is clear that enhanc-
ing SVM parameters in melanoma cases will also provide great
convenience to dermatologists in diagnosis process.
When Tables 7 and 8 are analyzed, especially considering the
accuracy values, it can be seen that the most successful results
are obtained with the proposed algorithm in the both datasets.
According to the Tables 7 and 8 we can easily assume that our pro-
posed algorithm is ﬂexibly apply on any dermoscopy images data-
sets for melanoma detection from images which have different
sizes and shapes. Hereby, we can offer users to not only a robust
but also a ﬂexible CAD system for detection of melanoma. When
the obtained results are compared, it is observed that the proposed
method is promising, especially according to deep learning
methods.
Fig. 9. Graphics of a) SVM (Scikit-learn) confusion matrix, b) SVM (Scikit-learn)
ROC curve, c) SVM (ThunderSVM) confusion matrix, d) SVM (ThunderSVM) ROC
curve, e) hybSVM confusion matrix, f) hybSVM ROC curve.
Sümeyya
_
Ilkin, Tug
˘rul Hakan Gençtürk, F. Kaya Gülag
˘ız et al. Engineering Science and Technology, an International Journal xxx (xxxx) xxx
10
In the light of all the tests and calculations, based on the meth-
ods and results in the studies mentioned above, it can be seen that
melanoma detection can be performed more successfully and fas-
ter with the proposed hybSVM algorithm. Based on the obtained
metric results and ROC curve analyses, it was demonstrated that
our study yielded successful results in the classiﬁcation of mela-
noma areas on skin images.
Although hybSVM has successful results, it has some limita-
tions. These are listed below.
Noisy images (images with hairs, discoloration, etc.) reduce the
performance of the proposed algorithm.
In cases where the wound part reaches the limits of the image,
classiﬁer effects of obtained features are decreasing within the
model.
The existing methods used to eliminate the noise (hairs, discol-
oration, etc.) in the images are not sufﬁcient. New methods are
needed to be developed in this subject in order to reduce the
impact of noise on the performance of the study. As a solution to
this problem, it is planned to develop new image enhancement
techniques in the future. In addition, while extracting the asymme-
try feature in the images, the lesion area of the image is divided
into two and handled symmetrically. While the lesion part is in
the middle of the image, it does not a problem for feature extrac-
tion process. However, it causes a general decrease in feature’s val-
ues when the lesion part comes to the edges. This situation
affecting the features also affects the accuracy of the classiﬁer. In
order to solve this problem, hardwares that will receive images
in a certain standard is needed. Especially considering today’s tech-
nological opportunities, it is seen that dermatoscope devices solve
this problem to a great extent. For this reason, the effect of this lim-
itation will be largely eliminated in professional images taken over
up to date dermatoscope devices.
6. Conclusions
Melanoma is one of deadly skin cancer types; and early detec-
tion of melanoma is vital. According to studies, early detection of
skin cancers increases the chances of treatment. Hence the diagno-
sis of this cancer type by using CAD systems is becoming more
common among the researchers.
For this purpose, in this study, melanoma skin lesions were
detected on dermoscopy images. In the proposed paper, we used
FSLIC clustering method. K center values of 4, 8, 16 and 32 were
used to determine the most successful k center value for segmen-
tation. The ROI images obtained after the segmentation were com-
pared with the ROI images from the data set and the accuracy,
recall, f-measure, Dice, Jac, Mcc metrics were calculated. As a result
of these calculations, it was observed that the most successful seg-
mentation was FSLIC with a k-centre value of 4. After that, 4 differ-
ent machine learning techniques, namely the RC, NBC, DT, SVM
classiﬁers were used to determine the best method for the detec-
tion of malign skin lesions. The SVM algorithm which was chosen
within the scope of the study was enhanced by using the bacterial
colony algorithm, to develop a hybrid classiﬁcation algorithm
which was named as hybSVM. The hybSVM algorithm, which uses
the heuristic optimization algorithm and operates in 2 phases, was
used to determine the gamma and C value with the optimum cost
and the operating cost of the SVM algorithm was improved in this
manner.
Table 6
Performance test results of hybSVM from ISIC dataset and PH2 dataset.
Dataset Time (sec.) MSE (%) Recall (%) Precision (%) F-Measure (%) Accuracy (%) AUC (%)
ISIC 26.50 15.81 97.94 96.95 97.44 97.56 98.00
PH2 11.9 15.81 93.75 100 96.77 97.50 97.00
Table 7
Comparison results of 10 cross fold validation with State-of-art.
Dataset Paper Methods Sensitivity (%) Speciﬁcity (%) Average Accuracy (%)
ISIC Proposed hybSVM 96.50 93.90 95.20
Murugan et al. [18] SVM 91.15 87.71 89.43
PH2 Proposed hybSVM 96.00 94.50 95.25
Alsaeed [19] SVM 90.00 96.00 92.6
Astorino et al. [22] MIL 95.24 84.53 90.63
Table 8
Comparison of results with State-of-art.
Dataset Paper Methods Sensitivity (%) Speciﬁcity (%) Precision (%) F-Measure (%) Accuracy (%) AUC
ISIC Proposed hybSVM 97.94 97.07 96.95 97.44 97.56 98.00
Esteva et al. [11] CNN-PA 72.10 94.00
Yue et al. [12] SVM 52.00 82.40 61.60 84.40 77.90
Pham et al. [14] SVM 54.70 95.00 71.50 87.20 87.40
Li and Shen [15] CNN 49.00 96.10 72.90 85.70 91.20
Hosny et al. [20] DCNN 88.47 93.00 – 95.91 –
Al-masni et al. [52] FrCN 85.40 96.69 – 94.03 –
Ünver and Ayan [54] YOLO v3 90.82 92.68 93.39
PH2 Proposed hybSVM 93.75 100 100 96.77 97.5 97.00
Suro
´wka and Ogorzalek [17] C-SVM (polynomial) 94.00
Arora et al. [23] Quadratic SVM 100 60 85.7 76.00
Al-masni et al. [52] FrCN 93.72 95.65 – 95.08 –
Kumar et al. [55] DE-ANN 96.9 –
Barros et al. [56] ANNM 87.5 90.9 – 90.0 –
Sümeyya
_
Ilkin, Tug
˘rul Hakan Gençtürk, F. Kaya Gülag
˘ız et al. Engineering Science and Technology, an International Journal xxx (xxxx) xxx
11
The study was tested with 1000 images in different sizes and
resolutions from the open access dataset generated as a part of
the International Skin Imaging Collaboration (ISIC) Melanoma Pro-
ject and 200 images from the open access PH2 dataset generated as
a part of the ADDI Project. The effectiveness and performance of
the study was assessed by calculating the time, accuracy, MSE, pre-
cision, recall, f-measure metrics and through the results obtained
by using the ROC curve. As a result of the comparison, it was con-
cluded that the best metric values and operation time was
achieved by hybSVM, which was developed within the scope of
the proposed algorithm, with Time = 26.5 sec, MSE = 0.158, Preci-
sion = 0.969, Recall = 0.979, F-Measure = 0.974, Accuracy = 0.975,
AUC = 0.98 for ISIC dataset and with Time = 11.9 sec,
MSE = 0.158, Precision = 0.937, Recall = 1.000, F-
Measure = 0.967, Accuracy = 0.975, AUC = 0.97 for PH2 dataset.
The effectiveness of the study was evaluated in comparison with
other studies in the literature and it was seen that the best AUC
value was achieved by our study. In addition, to verify the effec-
tiveness of the study, a State-of-art comparison was also per-
formed by using 10 cross validation. In the result of the
comparisons, it was concluded that the most successful results in
the detection of melanoma skin lesions were achieved through
the hybSVM algorithm proposed within the scope of the study.
The proposed algorithm presented in this paper shows us that
CAD systems have great potential for dermatologists as a reliable,
fast and more accurate diagnostic tool for fatal skin cancer types,
such as malign melanoma. Recently for the purpose of skin condi-
tion detection or condition tracking many applications developed
such as SkinVision [60], MoleScope, Miiskin, MoleMapper and
UMSkinCheck etc. [61]. The proposed hybSVM algorithm can be
easily integrated through a web service etc. to these applications
and it can detect melanoma highly accurately. At the same time,
the proposed algorithm is ready to use for dermatologists via a
web interface. Also, in the future, a new dataset which will be cre-
ated from melanoma images that stored on servers by using this
web interface, can be made available to researchers with the per-
mission of the patients/dermatologist.
Declaration of Competing Interest
The authors declare that they have no known competing ﬁnan-
cial interests or personal relationships that could have appeared
to inﬂuence the work reported in this paper.
Acknowledgments
We thank Huseyin Kalcik who is a member of the Embedded
Systems Laboratory at Kocaeli University, Computer Engineering
Department for his contribution to this study.
References
[1] ‘‘Memorial Antalya Hastanesi. Cilt (Deri) Kanseri Belirtileri, Tedavisi ve
Korunma Yöntemleri,” https://www.memorial.com.tr/saglik-rehberleri/cilt-
deri-kanseri-belirtileri-tedavisi-ve-korunma-yontemleri/, (visited: 2019-09-
15).
[2] H. Ganster, P. Pinz, R. Rohrer, E. Wildling, M. Binder, H. Kittler, Automated
melanoma recognition, IEEE Trans. Med. Imaging 20 (3) (2001) 233–239.
[3] S. Conoci, F. Rundo, S. Petralta, S. Battiato, Advanced skin lesion discrimination
pipeline for early melanoma cancer diagnosis towards PoC devices, in: 2017
European Conference on Circuit Theory and Design (ECCTD), 2017, pp. 1–4.
[4] N. Raut, A. Shah, S. Vira, H. Sampat, A study on different techniques for skin
cancer detection, Int. Res. J. Eng. Technol. 5 (9) (2018) 613–617.
[5] B. Mrunal, K. Ankita, S. Rutuja, W. Rutuja, Survey on machine learning
approach for the detection of melanoma using dermoscopic images, Int. J.
Emerg. Technol. Comput. Sci. 3 (4) (2018) 102–106.
[6] E.V. Carrera, D. Ron-Dominguez, A computer aided diagnosis system for skin
cancer detection, Int. Conf. Technol. Trends (2018) 553–563.
[7] A.B. Katapadi, M.E. Celebi, S.C. Trotter, M.N. Gurcan, Evolving strategies for the
development and evaluation of a computerised melanoma image analysis
system, Comput. Methods Biomech. Biomed. Eng.: Imaging Visualization 6 (4)
(2018) 465–472, https://doi.org/10.1080/21681163.2016.1277785.
[8] N. Hameed, A. Ruskin, K.A. Hassan, M.A. Hossain, A comprehensive survey on
image-based computer aided diagnosis systems for skin cancer, in: 2016 10th
International Conference on Software, Knowledge, Information Management
\& Applications (SKIMA), 2016, pp. 205–214.
[9] M.A. Marchetti, N.C.F. Codella, S.W. Dusza, A. David, B.H. Gutman, A. Kalloo, N.
Mishra, et al., Results of the 2016 international skin imaging collaboration
international symposium on biomedical imaging challenge: comparison of the
accuracy of computer algorithms to dermatologists for the diagnosis of
melanoma from dermoscopic images, J. Am. Acad. Dermatol. 78 (2) (2018)
270–277.
[10] Z. Waheed, A. Waheed, M. Zafar, F. Riaz, An efﬁcient machine learning
approach for the detection of melanoma using dermoscopic images, in: 2017
International Conference on Communication, Computing and Digital Systems
(C-CODE), 2017, pp. 316–319.
[11] A. Esteva, B. Kuprel, R.A. Novoa, J. Ko, S. Swetter, H. Blau, et al., Dermatologist-
level classiﬁcation of skin cancer with deep neural networks, Nature 542
(7639) (2017) 115–118.
[12] L. Yu, H. Chen, Q. Dou, J. Qin, P.A. Heng, Automated melanoma recognition in
dermoscopy images via very deep residual networks, IEEE Trans. Med. Imaging
36 (4) (2017) 994–1004.
[13] S. Mustafa, A. Kimura, A SVM-based diagnosis of melanoma using only useful
image features, in: International Workshop on Advanced Image Technology
(IWAIT), 2018, pp. 1–4.
[14] T.C. Pham, C.M. Luong, M. Visani, V.D. Hoang, Deep CNN and data
augmentation for skin lesion classiﬁcation, in: Asian Conference on
Intelligent Information and Database Systems, 2018, pp. 573–582.
[15] Y. Li, L. Shen, Skin lesion analysis towards melanoma detection using deep
learning network, Sensors 18 (2) (2018) 556.
[16] R.D. Seeja, A. Suresh, Deep learning based skin lesion segmentation and
classiﬁcation of melanoma using support vector machine (SVM), Asian Pac. J.
Cancer Prev. 20 (5) (2019) 1555–1561.
[17] G. Surówka, M. Ogorzalek, Resolution invariant wavelet features of melanoma
studied by SVM classiﬁers, PLoS One 14 (2) (2019), https://doi.org/10.1371/
journal.pone.0211318.
[18] A. Murugan, S.A.H. Nair, K.P.S. Kumar, Detection of skin cancer using SVM,
random forest and kNN classiﬁers, J. Med. Syst. 43 (8) (2019) 269.
[19] A.A.D. Alsaeed, On the development of a skin cancer computer aided diagnosis
system using support vector machine, Biosci. Biotechnol. Res. Commun. 12 (2)
(2019) 297–308.
[20] K.M. Hosny, M.A. Kassem, M.M. Foaud, Classiﬁcation of skin lesions using
transfer learning and augmentation with Alex-net, PLoS One 14 (2019) 5.
[21] S. Nasiri, J. Helsper, M. Jung, M.P.Y. Fathi, DePicT Melanoma Deep-CLASS: a
deep convolutional neural networks approach to classify skin lesion images,
BMC Bioinf. 21 (84) (2020) 1–13.
[22] A. Astorino, A. Fuduli, P. Veltri, E. Vocaturo, Melanoma detection by means of
multiple instance learning, Interdiscip. Sci.: Comput. Life Sci. 12 (1) (2020) 24–
31.
[23] G. Arora, A.K. Dubey, Z.A. Jaffery, R. Alvaro, Bag of feature and support vector
machine based early diagnosis of skin cancer, Neural Comput. Appl. (2020),
https://doi.org/10.1007/s00521-020-05212-y.
[24] G.N. Balaji, T.S. Subashini, N. Chidambaram, Detection and diagnosis of dilated
cardiomyopathy and hypertrophic cardiomyopathy using image processing
techniques, Eng. Sci. Technol., Int. J. 19 (4) (2016) 1871–1880.
[25] D. Połap, A. Winnicka, K. Serwata, K. Ke˛ sik, M. Woz
´niak, An intelligent system
for monitoring skin diseases, Sensors 18 (8) (2018) 2552.
[26] H. Zunair, A.B. Hamza, Melanoma detection using adversarial training and
deep transfer learning, Phys. Med. Biol. 65 (13) (2020) 135005.
[27] M. Woz
´niak, D. Połap, Bio-inspired methods modeled for respiratory disease
detection from medical images, Swarm Evol. Comput. 41 (2018) 69–96.
[28] H.A. Akkar, S.A. Salman, Detection of biomedical images by using bio-inspired
artiﬁcial intelligent, Eng. Technol. J. 38 (2 part (A) Engineering) (2020) 255–
264.
[29] L. Wang, P.C. Pedersen, E. Agu, D.M. Strong, B. Tulu, Area determination of
diabetic foot ulcer images using a cascaded two-stage SVM-based
classiﬁcation, IEEE Trans. Biomed. Eng. 64 (9) (2017) 2098–2109.
[30] A. _
Ince, S. Bozkurt, B. Bayram, Rasat Uydu Goruntusunden Nesne Tabanlı Kıyı
Cizgisi Cıkartma: Sapanca Golu Ornegi. TMMOB Harita ve Kadastro Muhendisleri
Odasi, 16, Turkiye Harita Bilimsel ve Teknik Kurultayi. (2017).
[31] U. Kose, Zeki Optimizasyon Tabanli Destek Vektor Makineleri ile Diyabet
Teshisi, Politeknik Dergisi. 22 (3) (2019) 557–566.
[32] A. Kim. Fast SLIC. https://github.com/Algy/fast-slic, (visited: 2019-10-30).
[33] D.T. Pham, S.S. Dimov, C.D. Nguyen, Selection of K in K-means clustering, Proc.
Inst. Mech. Eng., Part C: J. Mech. Eng. Sci. 219 (1) (2005) 103–119.
[34] E. Dincer, N. Duru, Girtlak Kanseri Ameliyat Verilerinin K-means Yontemiyle
Analizi, Turkiye Bilisim Vakﬁ Bilgisayar Bilimleri ve Muhendisligi Dergisi. 2
(2006) 1.
[35] Z. Cebeci, F. Yildiz, G.T. Kayaalp, K-ortalamalar kumelemesinde optimum K
degeri secilmesi. 2, Ulusal Yonetim Bilisim Sistemleri Kongresi. (2015) 231–242.
[36] R. Mukherjee, D.D. Manohar, D.K. Das, A. Achar, A. Mitra, C. Chakraborty,
Automated tissue classiﬁcation framework for reproducible chronic wound
assessment, Biomed Res. Int. 2014 (2014) 1–9, https://doi.org/10.1155/2014/
851582.
[37] T.Y. Tan, L. Zhang, M. Jiang, An intelligent decision support system for skin
cancer detection from dermoscopic images, in: 2016 12th International
Sümeyya
_
Ilkin, Tug
˘rul Hakan Gençtürk, F. Kaya Gülag
˘ız et al. Engineering Science and Technology, an International Journal xxx (xxxx) xxx
12
Conference on Natural Computation, Fuzzy Systems and Knowledge Discovery
(ICNC-FSKD), pp. 2194–2199.
[38] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M.
Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, Scikit-learn:
machine learning in Python, J. Mach. Learn. Res. 12 (2011) 2825–2830.
[39] Z. Wen, J. Shi, Q. Li, B. He, J. Chen, ThunderSVM: a fast SVM library on GPUs and
CPUs, J. Mach. Learn. Res. 19 (1) (2018) 797–801.
[40] Z. Aydın, Performance analysis of machine learning and bioinformatics
applications on high performance computing systems, Akademik Platform
Mühendislik ve Fen Bilimleri Dergisi 8 (1) (2020) 1–14, https://doi.org/
10.21541/apjes.547016.
[41] T.K. Ho, The random subspace method for constructing decision forests, IEEE
Trans. Pattern Anal. Mach. Intell. 20 (8) (1998) 832–844.
[42] M.M. Saritas, A. Yasar, Performance analysis of ANN and Naive Bayes
classiﬁcation algorithm for data classiﬁcation, Int. J. Intelligent Syst. Appl.
Eng. 7 (2) (2019) 88–91.
[43] H. Zhang, L. Jiang, L. Yu, Class-speciﬁc attribute value weighting for Naive
Bayes, Inf. Sci. (N. Y.) 508 (2020) 260–274.
[44] R.D.H. Devi, M.I. Devi, Outlier detection algorithm combined with decision tree
classiﬁer for early diagnosis of breast cancer, Int. J. Adv. Eng. Technol. 7 (2)
(2016) 93–98.
[45] ‘‘RBF SVM parameters,” https://scikit-learn.org/stable/auto_examples/
svm/plot_rbf_parameters.html#:~:text=The%20gamma%20parameters%20can
%20be,of%20the%20decision%20function’s%20margin, (visited: 2021-01-10).
[46] T.A. Gomes, R.B. Prudêncio, C. Soares, A.L. Rossi, A. Carvalho, Combining meta-
learning and search techniques to select parameters for support vector
machines, Neurocomputing 75 (1) (2012) 3–13.
[47] B. Niu, H. Wang, Bacterial colony optimization, Discrete Dyn. Nat. Soc. (2012).
[48] R. Kalyani, P.D. Sathya, V.P. Sakthivel, Trading strategies for image segmentation
using multilevel thresholding aided with minimum cross entropy, Eng. Sci.
Technol., Int. J. (2020), https://doi.org/10.1016/j.jestch.2020.07.007.
[49] F.F.X. Vasconcelos, A.G. Medeiros, S.A. Peixoto, P.P. Reboucas Filho, Automatic
skin lesions segmentation based on a new morphological approach via
geodesic active contour, Cognit. Syst. Res. 55 (2019) 44–59.
[50] N. Codella, V. Rotemberg, P. Tschandl, M.E. Celebi, S. Dusza, D. Gutman, B.
Helba et al. Skin lesion analysis toward melanoma detection 2018: A challenge
hosted by the international skin imaging collaboration (ISIC). arXiv preprint
arXiv:1902.03368. 2019.
[51] Y. Yuan, M. Chao, Y. Lo. Automatic skin lesion segmentation with fully
convolutional-deconvolutional networks. arXiv preprint arXiv:1703.05165.
2017.
[52] M.A. Al-Masni, M.A. Al-Antari, M.T. Choi, S.M. Han, T.S. Kim, Skin lesion
segmentation in dermoscopy images via deep full resolution convolutional
networks, Comput. Methods Programs Biomed. 162 (2018) 221–231.
[53] N. Sarang. Understanding AUC - ROC Curve. https://towardsdatascience.com/
understanding-auc-roc-curve-68b2303cc9c5, (visited: 2019-10-15).
[54] H.M. Ünver, E. Ayan, Skin lesion segmentation in dermoscopic images with
combination of YOLO and grabcut algorithm, Diagnostics 9 (3) (2019) 72.
[55] M. Kumar, M. Alshehri, R. AlGhamdi, P. Sharma, V. Deep, A DE-ANN inspired
skin cancer detection approach using fuzzy C-means clustering, Mobile
Networks Appl. 25 (2020) 1319–1329, https://doi.org/10.1007/s11036-020-
01550-2.
[56] W.K. Barros, D.S. Morais, F.F. Lopes, M.F. Torquato, R.D.M. Barbosa, M.A.
Fernandes, Proposal of the CAD system for melanoma detection using
reconﬁgurable computing, Sensors 20 (11) (2020) 3168.
[57] N. Mohan, K.P. Soman, R. Vinayakumar, Deep power: deep learning
architectures for power quality disturbances classiﬁcation, in: International
Conference on Technological Advancements in Power and Energy (TAP Energy)
, 2017, pp. 1–6, https://doi.org/10.1109/TAPENERGY.2017.8397249.
[58] S. Amiri, S. Salimzadeh, A.S.Z. Belloum, A survey of scalable deep learning
frameworks, in: 2019 15th International Conference on eScience (eScience),
2019, pp. 650–651, https://doi.org/10.1109/eScience.2019.00102.
[59] M. Woz
´niak, D. Połap, Intelligent home systems for ubiquitous user support by
using neural networks and rule-based approach, IEEE Trans. Ind. Inf. 16 (4)
(2019) 2651–2658.
[60] J.J. Deeks, J. Dinnes, H.C. Williams, Sensitivity and Speciﬁcity of SkinVision are
likely to have been overestimated, J. Eur. Acad. Dermatol. Venereol. 34 (2020)
e582–e583, https://doi.org/10.1111/jdv.16382.
[61] A.T. Young, N.B. Vora, J. Cortez, A. Tam, Y. Yeniay, L. Aﬁﬁ, D. Yan, A. Nosrati, A.
Wong, A. Johal, M.L. Wei, The role of technology in melanoma screening and
diagnosis, Pigm. Cell Melanoma Res. (2020) 1–13, https://doi.org/10.1111/
pcmr.12907.
Sümeyya
_
Ilkin, Tug
˘rul Hakan Gençtürk, F. Kaya Gülag
˘ız et al. Engineering Science and Technology, an International Journal xxx (xxxx) xxx
13
... This MLP approach is more efficient and reliable than traditional methods, but it comes at a high cost and requires an importable detector system. The SVM with bacterial colony optimization (BCO) algorithm was introduced byİlkin et al. [11]. The PH2 and ISIC datasets with10 cross-fold validation are used to examine the evaluation of the SVM-BCO algorithm for malignant melanoma classification.The processing time was 11.9 s and the AUC was 98%, but the feature dimensionality was higher. ...
Article
Full-text available
The most aggressive and malignant type of skin cancer is melanoma. When input data is in the form of images, image processing plays a vital role to detect and classifying cancer in the human body. Existing research discovered many weaknesses in complex data models, such as higher feature dimensionality, which required more data for training, resulting in lower detection accuracy, higher computational difficulties, portability, and processing time. Hence, we introduced fractal neural network-based galactic swarm optimization (FNN–GSO) algorithm for the detection of malignant melanoma such as superficial spreading, nodular, and lentigo malignant melanoma. The main aim of this work is to apply a deep learning technique to classify skin lesions for effective treatment and prognostication instead of the gold standard excision biopsy, which is currently used to diagnose this condition. Expert analysis, time consumption, and expensive processing associated with malignant melanoma classification and prediction are minimized in this manner. The proposed work involves four major components such as pre-processing, segmentation, feature extraction, and classification. The raw input images are pre-processed thereby the noise removal and contrast level enhancement are carried out. An adaptive watershed segmentation algorithm performs malignant melanoma segmentation. Following that, image features such as are extracted correctly using the DWT–GLCM feature extraction model. Finally, the malignant melanoma classifications are performed using the FNN–GSO algorithm. The proposed method's performance is evaluated using MATLAB software and different evaluation parameters. The proposed FNN–GSO algorithm demonstrates better classification results than other existing methods.
... In this experiment, precious, recall, false positive rate (FPR), and F-measure are selected as evaluation metrics of the proposed method, which are calculated as follows [22,23]: ...
Article
Full-text available
A sports-assisted education method based on a support vector machine (SVM) is proposed to address the problem of complex and variable sports actions leading to easy ghosting of target detection and high dimensionality of feature extraction, which reduces the low accuracy of sports action recognition. The ViBe target detection algorithm is improved by using Wronskian function and the “4-linked algorithm” seed filling algorithm, which effectively solves the ghosting problem and obtains clearer human sports targets. By using the genetic algorithm to fuse the eight-star model with sports action features extracted by the Zernike moment, redundant features are reduced and differentiability between different classes is ensured. Sports action classification was achieved by using a one-to-one construction of an SVM classifier. The results show that the proposed method can effectively recognize sports movements with an average recognition accuracy of more than 96%, which can assist physical education and has a certain practical application value.
... (2) Assign flight direction and distance to all fruit flies and olfactory search is utilized to update [23] X i � X axis + rand(FR), ...
Article
Full-text available
To improve the ability of market to avoid and prevent credit risk and strengthen the awareness of market risk early warning, SMOTE is used to process the unbalanced sample, and fruit fly optimization algorithm (FOA) is utilized to optimize the parameters of support vector machine (SVM), and thus an improved SVM market risk early warning model is proposed. The simulation results show that the proposed model has excellent stability and generalization ability, and it can predict market credit risk accurately. Compared with the prediction model based on FOA-SMOTE-BP and FOA-SMOTE-Logit, the proposed model performs better on the indicators of G value, F value, and AUC value, which provides a reference for market credit risk prediction.
... As convex quadratic programming can only be solved by a global optimal solution, which makes the process of solution simple, the global optimal solution can be derived by calculating the extrema. In solving convex quadratic programming, a combination of structural and empirical risk needs to be considered, and we can get the following equation after the Lagrangian function is introduced into equations (3)-(4) based on the Lagrangian duality [22]: ...
Article
Full-text available
This paper presents a settlement prediction method based on PSO optimized SVM for improving the accuracy of foundation pit settlement prediction. Firstly, the method uses the SA algorithm to improve the traditional PSO algorithm, and thus, the overall optimization-seeking ability of the PSO algorithm is improved. Secondly, the improved PSO algorithm is used to train the SVM algorithm. Finally, the optimal SVM model is obtained, and the trained model is used in foundation pit settlement prediction. The results suggest that the settling results obtained from the optimized model are closer to the actual values and also more advantageous in indicators such as RMSE. The fitting value R2 = 0.9641, which is greater, indicates a better fitting effect. Thus, it is indicated that the improvement method is feasible.
Article
Full-text available
Skin cancer affects the lives of millions of people every year, as it is considered the most popular form of cancer. In the USA alone, approximately three and a half million people are diagnosed with skin cancer annually. The survival rate diminishes steeply as the skin cancer progresses. Despite this, it is an expensive and difficult procedure to discover this cancer type in the early stages. In this study, a threshold-based automatic approach for skin cancer detection, classification, and segmentation utilizing a meta-heuristic optimizer named sparrow search algorithm (SpaSA) is proposed. Five U-Net models (i.e., U-Net, U-Net++, Attention U-Net, V-net, and Swin U-Net) with different configurations are utilized to perform the segmentation process. Besides this, the meta-heuristic SpaSA optimizer is used to perform the optimization of the hyperparameters using eight pre-trained CNN models (i.e., VGG16, VGG19, MobileNet, MobileNetV2, MobileNetV3Large, MobileNetV3Small, NASNetMobile, and NASNetLarge). The dataset is gathered from five public sources in which two types of datasets are generated (i.e., 2-classes and 10-classes). For the segmentation, concerning the “skin cancer segmentation and classification” dataset, the best reported scores by U-Net++ with DenseNet201 as a backbone architecture are 0.104, 94.16%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$94.16\%$$\end{document}, 91.39%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$91.39\%$$\end{document}, 99.03%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$99.03\%$$\end{document}, 96.08%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$96.08\%$$\end{document}, 96.41%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$96.41\%$$\end{document}, 77.19%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$77.19\%$$\end{document}, 75.47%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$75.47\%$$\end{document} in terms of loss, accuracy, F1-score, AUC, IoU, dice, hinge, and squared hinge, respectively, while for the “PH2” dataset, the best reported scores by the Attention U-Net with DenseNet201 as backbone architecture are 0.137, 94.75%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$94.75\%$$\end{document}, 92.65%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$92.65\%$$\end{document}, 92.56%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$92.56\%$$\end{document}, 92.74%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$92.74\%$$\end{document}, 96.20%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$96.20\%$$\end{document}, 86.30%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$86.30\%$$\end{document}, 92.65%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$92.65\%$$\end{document}, 69.28%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$69.28\%$$\end{document}, and 68.04%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$68.04\%$$\end{document} in terms of loss, accuracy, F1-score, precision, sensitivity, specificity, IoU, dice, hinge, and squared hinge, respectively. For the “ISIC 2019 and 2020 Melanoma” dataset, the best reported overall accuracy from the applied CNN experiments is 98.27%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$98.27\%$$\end{document} by the MobileNet pre-trained model. Similarly, for the “Melanoma Classification (HAM10K)” dataset, the best reported overall accuracy from the applied CNN experiments is 98.83%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$98.83\%$$\end{document} by the MobileNet pre-trained model. For the “skin diseases image” dataset, the best reported overall accuracy from the applied CNN experiments is 85.87%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$85.87\%$$\end{document} by the MobileNetV2 pre-trained model. After computing the results, the suggested approach is compared with 13 related studies.
Article
The dealers are responsible for connecting customers to manufacturers in the automotive industry. An effective relationship between the dealers and the manufacturers can lead to customer satisfaction and loyalty. This research provides an approach to evaluate factors affecting the dealers-manufacturer relationship and develop a solution for predicting how manufacturers can continue or terminate the cooperation with their dealers in the after-sales service network, which is less discussed in the literature. We used predictive analytics tools and classification methods to predict dealers’ cooperation as a dependent variable in two classes based on a 2-year real-life collected data. Our results revealed that dealer performance violations, changing dealer principles, and the services revenue have the most effect on the cooperation of dealers in the automotive after-sales service network. Also, this research proposed a dealer management solution for manufacturers to show them the correct decisions for the dealers’ cooperation at the right time. This solution can be used in real-life cases in after-sales service networks to prevent using redundant resources, improve the weak performance of dealers, avoid unnecessary termination of cooperation, avoid the costs of attracting new dealers, and decrease customer dissatisfaction.
Article
Full-text available
Multilevel thresholding (MLT) is one of the most widely used methods in image segmentation. However, the exhaustive search method is computationally time consuming for selecting the optimal thresholds. Consequently, heuristic algorithms are extensively used to reduce the complexity of the MLT problem. In this paper, an efficient Exchange Market Algorithm (EMA) is proposed to segment images using minimum cross entropy thresholding method. In the EMA, a market risk variable is used to balance the exploration and exploitation capabilities of the algorithm. Moreover, the local search capability is strengthened by the search and absorbent operators of EMA. Meanwhile, the most competent shareholders of EMA retain their best rank without undergoing any changes in their shares. These help in reducing the computational time. The proposed EMA based MLT is tested on benchmark and brain images with different threshold levels. Additionally, EMA approach is compared with other well-known algorithms such as, genetic algorithm, particle swarm optimization, bacterial foraging algorithm, firefly algorithm, honey bee mating optimization and teaching–learning based optimization. The experimental results show that the proposed EMA approach provides better outcomes than other algorithms.
Article
Full-text available
Skin cancer is one of the diseases which lead to death if not detected at an early stage. Computer-aided detection and diagnosis systems are designed for its early diagnosis which may prevent biopsy and use of dermoscopic tools. Numerous researches have considered this problem and achieved good results. In automatic diagnosis of skin cancer through computer-aided system, feature extraction and reduction plays an important role. The purpose of this research is to develop computer-aided detection and diagnosis systems for classifying a lesion into cancer or non-cancer owing to the usage of precise feature extraction technique. This paper proposed the fusion of bag-of-feature method with speeded up robust features for feature extraction and quadratic support vector machine for classification. The proposed method shows the accuracy of 85.7%, sensitivity of 100%, specificity of 60% and training time of 0.8507 s in classifying the lesion. The result and analysis of experiments are done on the PH2 dataset of skin cancer. Our method improves performance accuracy with an increase of 3% than other state-of-the-art methods.
Article
Full-text available
As per recent developments in medical science, the skin cancer is considered as one of the common type disease in human body. Although the presence of melanoma is viewed as a form of cancer, it is challenging to predict it. If melanoma or other skin diseases are identified in the early stages, prognosis can then be successfully achieved to cure them. For this, medical imaging science plays an essential role in detecting such types of skin lesions quickly and accurately. The application of our approaches is to improve skin cancer detection accuracy in medical imaging and further, can be automated using electronic devices such as mobile phones etc. In the proposed paper, an improved strategy to detect three type of skin cancers in early stages are suggested. The considered input is a skin lesion image which by using the proposed method, the system would classify it into cancerous or non-cancerous type of skin. The image segmentation is implemented using fuzzy C-means clustering to separate homogeneous image regions. The preprocessing is done using different filters to enhance the image attributes while the other features are assessed by implementing rgb color-space, Local Binary Pattern (LBP) and GLCM methods altogether. Further, for classification, artificial neural network (ANN) is trained using differential evolution (DE) algorithm. Various features are accurately estimated to achieve better results using skin cancer image datasets namely HAM10000 and PH2. The novelty of the work suggests that DE-ANN is best compared among other traditional classifiers in terms of detection accuracy as discussed in result section of this paper. The simulated result shows that the proposed technique effectually detects skin cancer and produces an accuracy of 97.4%. The results are highly accurate compare to other traditional approaches in the same domain.
Article
Full-text available
This work proposes dedicated hardware to real-time cancer detection using Field-Programmable Gate Arrays (FPGA). The presented hardware combines a Multilayer Perceptron (MLP) Artificial Neural Networks (ANN) with Digital Image Processing (DIP) techniques. The DIP techniques are used to extract the features from the analyzed skin, and the MLP classifies the lesion into melanoma or non-melanoma. The classification results are validated with an open-access database. Finally, analysis regarding execution time, hardware resources usage, and power consumption are performed. The results obtained through this analysis are then compared to an equivalent software implementation embedded in an ARM A9 microprocessor.
Article
Full-text available
Skin lesion datasets consist predominantly of normal samples with only a small percentage of abnormal ones, giving rise to the class imbalance problem. Also, skin lesion images are largely similar in overall appearance owing to the low inter-class variability. In this paper, we propose a two-stage framework for automatic classification of skin lesion images using adversarial training and transfer learning toward melanoma detection. In the first stage, we leverage the inter-class variation of the data distribution for the task of conditional image synthesis by learning the inter-class mapping and synthesizing under-represented class samples from the over-represented ones using unpaired image-to-image translation. In the second stage, we train a deep convolutional neural network for skin lesion classification using the original training set combined with the newly synthesized under-represented class samples. The training of this classifier is carried out by minimizing the focal loss function, which assists the model in learning from hard examples, while down-weighting the easy ones. Experiments conducted on a dermatology image benchmark demonstrate the superiority of our proposed approach over several standard baseline methods, achieving significant performance improvements. Interestingly, we show through feature visualization and analysis that our method leads to context based lesion assessment that can reach an expert dermatologist level.
Article
Full-text available
Udrea et al [1] reported 95.1% sensitivity and 78.3% specificity using the SkinVision app for risk assessment of skin lesions. These are likely to be overestimates due to the nature of the study design and flaws in sampling.
Article
Full-text available
Computer vision and image processing are extremely necessary for medical pictures analysis. During this paper, a method of Bio-inspired Artificial Intelligent (AI) optimization supported by an artificial neural network (ANN) has been widely used to detect pictures of skin carcinoma. A Moth Flame Optimization (MFO) is utilized to educate the artificial neural network (ANN). A different feature is an extract to train the classifier. The comparison has been formed with the projected sample and two Artificial Intelligent optimizations, primarily based on classifier especially with, ANN-ACO (ANN training with Ant Colony Optimization (ACO)) and ANN-PSO (training ANN with Particle Swarm Optimization (PSO)). The results were assessed using a variety of overall performance measurements to measure indicators such as Average Rate of Detection (ARD), Average Mean Square error (AMSTR) obtained from training, Average Mean Square error (AMSTE) obtained for testing the trained network, the Average Effective Processing Time (AEPT) in seconds, and the Average Effective Iteration Number (AEIN). Experimental results clearly show the superiority of the proposed (ANN-MFO) model with different features.
Article
Full-text available
Background: Melanoma results in the vast majority of skin cancer deaths during the last decades, even though this disease accounts for only one percent of all skin cancers' instances. The survival rates of melanoma from early to terminal stages is more than fifty percent. Therefore, having the right information at the right time by early detection with monitoring skin lesions to find potential problems is essential to surviving this type of cancer. Results: An approach to classify skin lesions using deep learning for early detection of melanoma in a case-based reasoning (CBR) system is proposed. This approach has been employed for retrieving new input images from the case base of the proposed system DePicT Melanoma Deep-CLASS to support users with more accurate recommendations relevant to their requested problem (e.g., image of affected area). The efficiency of our system has been verified by utilizing the ISIC Archive dataset in analysis of skin lesion classification as a benign and malignant melanoma. The kernel of DePicT Melanoma Deep-CLASS is built upon a convolutional neural network (CNN) composed of sixteen layers (excluding input and ouput layers), which can be recursively trained and learned. Our approach depicts an improved performance and accuracy in testing on the ISIC Archive dataset. Conclusions: Our methodology derived from a deep CNN, generates case representations for our case base to use in the retrieval process. Integration of this approach to DePicT Melanoma CLASS, significantly improving the efficiency of its image classification and the quality of the recommendation part of the system. The proposed method has been tested and validated on 1796 dermoscopy images. Analyzed results indicate that it is efficient on malignancy detection.
Article
Melanoma presents challenges for timely and accurate diagnosis. Expert panels have issued risk‐based screening guidelines, with recommended screening by visual inspection. To assess how recent technology can impact the risk/benefit considerations for melanoma screening, we comprehensively reviewed noninvasive visual‐based technologies. Dermoscopy increases lesional diagnostic accuracy for both dermatologists and primary care providers; total body photography and sequential digital dermoscopic imaging also increase diagnostic accuracy, are supported by automated lesion detection and tracking, and may be best suited to use by dermatologists for longitudinal follow‐up. Specialized imaging modalities using non‐visible light technology have unproven benefit over dermoscopy and can be limited by cost, access, and training requirements. Mobile apps facilitate image capture and lesion tracking. Teledermatology has good concordance with face‐to‐face consultation and increases access, with increased accuracy using dermoscopy. Deep learning models can surpass dermatologist accuracy, but their clinical utility has yet to be demonstrated. Technology‐aided diagnosis may change the calculus of screening; however well‐designed prospective trials are needed to assess the efficacy of these different technologies, alone and in combination to support refinement of guidelines for melanoma screening.