Journal of Medical Systems (2019) 43: 289

https://doi.org/10.1007/s10916-019-1413-3

IMAGE & SIGNAL PROCESSING

Region Extraction and Classification of Skin Cancer: A Heterogeneous

framework of Deep CNN Features Fusion and Reduction
Tanzila Saba 1 & Muhammad Attique Khan 2 & Amjad Rehman 3 & Souad Larabi Marie-Sainte 1

Received: 5 March 2019 / Accepted: 3 July 2019 / Published online: 20 July 2019
# Springer Science+Business Media, LLC, part of Springer Nature 2019

Abstract
Cancer is one of the leading causes of deaths in the last two decades. It is either diagnosed malignant or benign – depending upon the
severity of the infection and the current stage. The conventional methods require a detailed physical inspection by an expert derma-
tologist, which is time-consuming and imprecise. Therefore, several computer vision methods are introduced lately, which are cost-
effective and somewhat accurate. In this work, we propose a new automated approach for skin lesion detection and recognition using a
deep convolutional neural network (DCNN). The proposed cascaded design incorporates three fundamental steps including; a) contrast
enhancement through fast local Laplacian filtering (FlLpF) along HSV color transformation; b) lesion boundary extraction using color
CNN approach by following XOR operation; c) in-depth features extraction by applying transfer learning using Inception V3 model
prior to feature fusion using hamming distance (HD) approach. An entropy controlled feature selection method is also introduced for
the selection of the most discriminant features. The proposed method is tested on PH2 and ISIC 2017 datasets, whereas the recognition
phase is validated on PH2, ISBI 2016, and ISBI 2017 datasets. From the results, it is concluded that the proposed method outperforms
several existing methods and attained accuracy 98.4% on PH2 dataset, 95.1% on ISBI dataset and 94.8% on ISBI 2017 dataset.

Keywords Skin cancer . Augmentation . Contrast improvement . Boundary extraction . Deep learning . Features selection

Introduction diagnosed cancers cases and the incidence rate continues

to rise worldwide. The factors of increase rest in ozone
The studies of the World Health Organization (WHO) layer reduction which is responsible for the protection
assess that skin cancer has a share of one-third of all against ultraviolet radiations. Skin cancer has two types:
melanoma and non-melanoma. More than 87,000 mela-
nomas and over five million non-melanoma cases are
Highlights diagnosed each year only in the US [1]. Melanoma is
• This research presents a novel approach for microscopic skin lesion
boundary extraction and lesion recognition through Deep Neural Network.
the most dangerous type of skin cancer and begins in
• Distinction features are selected through clustering controlled entropy pigment cells. Approximately 75% of deaths related to
approach and classified through MLP. skin cancer occurred due to malignant lesions [2].
This article is part of the Topical Collection on Image & Signal Doctors in the clinics use some advanced and sophisti-
Processing cated treatment methods and tools, such as immunother-
apy and radiation therapy along with surgery clinical
* Amjad Rehman tools. The five-year survival ratio of advanced stage
[email protected] cases is still below than 15%, whereas the survival rate
for early-stage cases is over 95% [3]. This difference
Tanzila Saba
identifies that the patient survival rate highly depends
[email protected]
on the timely detection and treatment of melanoma.
1
College of Computer and Information Sciences, Prince Sultan Therefore, the primary goal is to diagnose melanoma
University, Riyadh 11586, Saudi Arabia in its early stages [4, 5].
2
Department of Computer Science and Engineering, HITEC In order to improve the melanoma diagnostic accuracy,
Universit, Museum Road, Taxila, Pakistan most of the dermatologists used a popular noninvasive
3
College of Computer and Information Sciences, Imam Mohammad imaging method known as dermoscopy. In dermoscopy,
Ibn Saud Islamic University, Riyadh, Saudi Arabia first, a gel is placed on the lesion area. Later, by using a
289 Page 2 of 19 J Med Syst (2019) 43: 289

magnifier tool, an illuminated and magnified image ac- detection and classification through enhancing color
quires which provides better visualization of the pigment convolutional network (ECCN) and fusion of fully con-
structure of the lesion [6]. It improves the diagnostic rate nected and average pool layers of inception v3 Deep
significantly compared to the examination with a naked CNN model. Our major contributions are:
eye. Moreover, it has been showed that dermoscopy has a
detection accuracy reaching approximately 75% [7]. The 1) In the preprocessing phase, an augmentation is performed
doctors used a few well-known diagnostics tools in the in the initial step for an equal length of all classes of
clinics such as ABCDE rule, seven-point checklist, optical cancer types. Then, enhancing the images through fast
techniques such as laser and few more. However, the local Laplacian filtering (FlLpF) along with HSV color
manual interpretation of dermoscopic images is time- transformation defined by (FlLpFaHSV) for the next
consuming and subject of dermatologist experience and process.
clinical training. Besides, an experienced dermatologist 2) The enhanced HSV transformed data is utilized for
may produce an error during the diagnostic process [8]. the segmentation process, in which the color labels
Therefore, a computerized based diagnostic and analysis are assigned (Pink and Yellow). The yellow color de-
methods are required will assist the dermatologist for fast picts healthy pixels, whereas the Pink color describes
detection and to improve the detection accuracy. The re- lesion pixels. Through color labels, the sample is split
searchers’ community from the field of image processing into 60% for training the deep CNN model and 40%
and machine learning (ML) put great efforts in the devel- for testing the segmentation effects. Additionally, the
opment of automated based melanoma diagnostic tools thresholding and morphological operations are uti-
referred to as computer-aided diagnosis (CAD) systems. lized for the retrieval of a binary image.
These methods overcome the issues of traditional 3) The deep CNN features are extracted through FC and AP
dermoscopy techniques [9]. layers of Inception V3 pre-trained model, then the fea-
The performance of ML in the area of medical image tures are mapped to the selected dermoscopy dataset
analysis has shown significant outputs in the detection, through fine-tuning. The extracted FC and AP layer fea-
diagnosis, and analysis of medical objects [10–13]. For tures are fused through Hamming distance.
example, skin lesions are complex and difficult to predict 4) The fused CNN features are divided into two clusters
precisely while using a simple design [14]. Through these based on the mean threshold value and the computed
ML designs, the objects like lesions are categorized into entropy value. Then, the posterior probability values
their relevant classes such as melanoma, basal cell carci- are computed for each cluster. After that, the initial
noma, and benign. These models are based on the feature probability values are updated until the error rate is
based ML design which is harder for a massive number of close to zero. This process almost reduces the features
images dataset [15]. Lately, the emergence of deep learn- number by 60%. Finally, concatenate both cluster fea-
ing in the area of ML has given a higher performance in tures after clustering controlled entropy (CcE) is per-
medical imaging. The DL based on the classification pro- formed and then apply one of the famous classifica-
cess is also known as image-based ML because the im- tion methods such as SVM, K-Nearest Neighbor, and
ages are directly put to the DL design as input. The deep neural network (NN).
learning methods are highly useful when the number of
training inputs is large [16]. The rest of the paper is organized into four main sections.
Section “Related work” presents the recent state of the art.
Problem statement and contributions Section “Proposed methodology” exhibits the proposed meth-
odology. Section “Experimental results and discussion” ad-
The computerized based melanoma detection and recogni- dresses the results and discussion. Finally, section
tion is still a very challenging task. Few of major general “Conclusion” concludes the research.
issues are similar colors of healthy and skin lesions, change
in lesion shapes, texture, change among normal and skin
lesion, and presence of a lesion on the boundary areas. Related work
Some other artifacts such as illumination, veins, hairs, air
bubbles, color calibration marks, and many more affect the In the last two decades, different computer-based systems
segmentation and reduce the feature extraction accuracy have been developed to solve the challenges related to
[17]. By using deep learning, the researchers mostly faced melanoma segmentation and recognition. Four primary
the problems of irrelevant features which reduce the recog- steps are performed in any computerized method such as
nition accuracy of the designed DL model. In this work, we i) preprocessing, ii) lesion boundary detection, iii) feature
design a new heterogeneous model for skin lesion extraction and selection, and iv) classification. The
J Med Syst (2019) 43: 289 Page 3 of 19 289

preprocessing step is performed to increase the quality of In the developed approach, intra and inter architectures
images and removes some artifacts like hairs, air bubbles, are merged for features abstraction levels and each archi-
illumination, and poor contrast which may complicate the tecture consists of multiple pre-trained CNN networks and
next step. As a matter of fact it is the main requirement. fine-tuned on given dermoscopy images. Finally, the ex-
Many approaches have been utilized for removing these tracted features through fine-tuning are fed to SVM for
challenges such as software-based hair removal, morpho- classification. The experiments are performed on
logical operation bases removal, filtering, contrast ISIC2017. The obtained results outperformed the top-
stretching, masking, and few more. Lee et al. [18] present- ranked techniques. Pedro et al. [30] extracted structural
ed a software-based hair removal method referred to as co-occurrences matrices (SCM) for textural features ex-
DullRazor using morphological operations. Kiani et al. traction from dermoscopic images. These extracted fea-
[19] improves the DullRazor method and make it capable tures have powerful discrimination compared to other tex-
of removing both thick and light hairs from dermoscopic tural features. The evaluation is performed on ISIC 2016
images. [20] proposed a method for illumination correc- and 2017 datasets and reached a specificity of more than
tion from lesion images by using polynomial curve fitting 90%. The extracted features are mostly classified through
and bilateral decomposition. Most recently, Nasir et al. [4, supervised learning algorithms based on their pros and
5] utilized texture and color information of lesion for con- cons. SVM classifier is the most commonly used in com-
trast enhancement. Lesion Segmentation is the second puterized lesion classification. Other widely used classi-
most important step for lesion segmentation. The prosper- fiers in the literature include decision trees, k-nearest
ity of these analysis systems is highly dependent on the neighbor, artificial neural networks (ANN), and linear dis-
reliability and robustness of the segmentation step. It is criminate analysis (LDA) [31, 32].
complicated to develop an accurate and efficient segmen-
tation algorithm by considering different sizes, shapes,
color, and texture of lesions as well as different types of
Proposed methodology
skins. To conquer these problems, different approaches
are used in theliterature. These approaches are based on
The proposed lesion segmentation and recognition system
region which use region growing techniques [21], cluster-
are based on the contrast-enhanced deep CNN design
ing methods which differentiate the lesion pixels from
which comprises three primary steps: (a) data augmenta-
healthy skin pixels and homogeneous areas [22],
tion and contrast-enhanced for lesion visibility, (b) color
thresholding techniques [22], and active contours ap-
value based lesion identification and lesion boundary ex-
proaches which iteratively detect the lesion border [23].
traction, and (c) DCNN features extraction, fusion, and
Feature extraction step is involved in the extraction of
selection for recognition. Flow architecture of the pro-
essential features which correctly describe the current
posed system is shown in Fig. 1.
problem. After that, the selection of efficient one is re-
quired for reducing the computational complexity and in-
crease the overall system accuracy. Color, shape, and tex- Dataset augmentation & contrast enhancement
ture features are the most widely used features for lesion
detection [24]. Sadeghi et al. [25] explain the importance In medical imaging, the problem of imbalance classes
of irregular patterns in the lesion for diagnosis process. affects the accuracy of the system. Therefore, it is essen-
However, these features do not perform well for a large tial to resolve the imbalance issue through the rotation,
number of training samples. These irregular patterns in- filtering, and a few other techniques [33]. Nowadays, in
clude color and texture features. Satheesha et al. [26] ex- medical imaging, skin cancer is a very hot topic in re-
tract shape (2D, 3D), texture, and color features of the search, but the availability of datasets in this area is a
lesion and select the most appropriate features by using challenge for researchers. The available datasets include
heuristic approach. The importance of color features in a vast difference between the number of images in each
the classification process is presented in [27]. A clustering class. For example, PH2 dataset includes a total of 200
based method is used to extract color features. Moreover, images divided into40images in melanoma class, 80im-
automated segmentation of skin lesions through CNN has ages in common nevi class, and 80 images in atypical
given a significant performance compared to recent well- nevi class. The melanoma class has half the number of
known techniques like thresholding, etc. Philip et al. [28] images containing in the other two classes which causes
described an encoder weighted based approach for lesion degradation of the recognition accuracy. In this work, we
segmentation. They utilized ResNet-34 approach along rotate the images into 4 different angles (450, 600, 1350,
with ISIC2017 dataset. Mahbod et al. [29] designed an 2700) for the principle of data augmentation. This process
ensemble CNN approach for skin lesion segmentation. is performed only in the training phase.
289 Page 4 of 19 J Med Syst (2019) 43: 289

Fig. 1 Proposed workflow of stomach infection segmentation and recognition

Following, the augmentation step, color enhancement (FlLpF) along with the HSV color transformation. This
is performed through the fast local Laplacian filtering approach processes through two steps. Initially, the

Fig. 2 Proposed FlLpFaHSV enhancement effects. The first row represents original images, second row shows the FILpF effects, and bottom row
demonstrate the effects after HSV transformation
J Med Syst (2019) 43: 289 Page 5 of 19 289

Fig. 3 XOR operation effects on sample images. a Enhanced image, b XOR operation, c selection of first channel after XOR operation, d selection of
lesion part for identifying their pixels, e histogram of first channel of XOR image, and f selection range of 110-460 for lesion pixels

FlLpF approach is implemented through the following denotes the laplacian pyramid (LP) which constructs co-
formulation: efficients, where each constructed coefficient is indepen-
Let have the input image I of dimensions 512 × 512 at dent to other coefficients. At the initial stage, the simple
position(x, y). Let φ denotes the results output image of pixel wise filtering is performed on the original imageI,
dimension 512 × 512 at position(x, y). Let {ξ L [φ x, y ]} then compute LP through a transformed image. Through

Fig. 4 Proposed segmentation effects through previous pixels color information. a Original image, b CNN labeled image, c thresholding, d refinement,
and e mapped RGB
289 Page 6 of 19 J Med Syst (2019) 43: 289

Fig. 5 Proposed boundary extraction effects. a Original image, b proposed segmented, c ground truth, and d boundary detection

LP, the laplacian coefficient ξLl[φx, y]is computed at level highlighted edges of the lesion parts are further improved. The
l and position(x, y). This process is repeated for all coef- mathematical formulation of HSV is represented by ξHSV
ficients [34]. andthe formulation is done through Eq. (1-3).
After computing the resultant ξLl[φx, y]image, the HSV
transformation is applied. Through HSV transformation, the ξH ¼ 600  ξH 0 ð1Þ

Fig. 6 Architecture of Inception V3 framework

J Med Syst (2019) 43: 289 Page 7 of 19 289

Fig. 7 Proposed HD based fusion representation

8 boundary extraction, it is essential to find out the lesion from

>
> ξG −ξB
>
> if M ¼ ξR the input image. In the literature, various types of techniques
>
< C
ξG −ξB are introduced for lesion segmentation and boundary extrac-
ξH 0 ¼ þ 2 if M ¼ ξG ð2Þ
>
> C tion. The most famous methods are thresholding based lesion
>
> ξ −ξ
>
: G B þ 4 if M ¼ ξB extraction [35], region growing based lesion segmentation
C [36], CNN based segmentation [37], and few more [38, 39].
8
< 0 ξV ¼ 0 From all, the deep learning based segmentation techniques are
ξS ¼ C ð3Þ well accurate in terms of automation and efficiency.
: Oterwise
ξV In this work, we proposed a new deep neural network
(DCNN) approach for skin lesion segmentation and later
Where, V = M, M = max(ξR, ξG, ξB), C is a constant value identification. In the proposed approach, initially, we di-
which is defined through the difference of the maximum and vide the dataset into training and testing. Both training
the minimum image pixels information as C = M − m, and and testing samples include both original and their ground
m = min(ξR, ξG, ξB). TheξR, ξG, and ξB represent the red, green, truth images. The stepwise formulation of this process is
and blue channels of the transformed image ξLl[φx, y]. The defined as follows:
sample resultant images are shown in Fig. 2, which are later Let, T1 denotes training and testing images set, T2 denotes
utilized for lesion identification. their corresponding labels set, then before further processing,
we perform XOR operation on the enhanced image ξHSV and
FlLpF image ξLl for range selection of lesion pixels. The XOR
Boundary extraction
operation is defined as:
   
The boundary extraction of skin lesion is an essential and
ξL l ðiÞ∪ ξHSV ð jÞ ¼ ξL l ðiÞ∩!ξHSV j ∪ !ξL l ðiÞ∩ξHSV ð jÞ ð4Þ
challenging task in the area of computer vision. In general,
the boundary extraction phase consists of two steps including    
lesion segmentation and identification. Before performing the ¼ ξlL ðiÞ∪ξHSV ð jÞ ∩ !ξlL ðiÞ∩!ξHSV ð jÞ ð5Þ
289 Page 8 of 19 J Med Syst (2019) 43: 289

Fig. 8 Proposed labeled results. a Original dermoscopy image, b proposed labeled image

Where ∪ denotes union operation, ∩ denotes intersection few of essential parameters are also defined like padding
operation, and ! explains not operation. The effects of XOR mode is manual, pooling size is 2 × 2, a number of max epochs
operation are shown in Fig. 3. In Fig. 3c, the first channel are 200, Min Batch Size is 64, and the learning rate is 0.2.
of XOR image is extracted and later select the most nearly After that, the activation is performed through the following
lesion pixel points. The probability is computed to find out loss function:
the exact lesion pixel values that later utilized for training 1
the CNN model. Through the probability value, it is clearly Fξ ¼ − ∑ ð6Þ
N i  M j i; j
showed that the pixels in range of 0-0.4 are best h       i
representing the lesion pixels, while pixels greater than this yi; j lnp wjξi; j þ 1−yi; j ln 1−p wjξi; j
range represent background or healthy regions, as shown
in Fig. 3f. The maximum pixel value is selected 0.4 using Where, yi, j denotes class labels ((yi, j) ∈ {pink, yellow})
trial and error method. and ξ i , j are actual pixels of the image. After
These lesion and healthy pixels are defined into two sepa- performing this activation function, anOtsu threshold
rate classes denoted by Δ1and Δ2, respectively. Then, load operation is performed and later refined through mor-
VGG19 pre-trained CNN model. Through fine-tuning, train phological operations such as closing, filling, and area
the model using lesion and healthy region pixel values. The removal [40] as shown in Fig. 4.
J Med Syst (2019) 43: 289 Page 9 of 19 289

Fig. 9 Samples dermoscopy images

Moreover, the boundary extraction is done through the Where, ξGR(i) denotes the ground truth image pixels and
comparison of the proposed segmented and the ground truth ξRG(j) denotes the proposed refined segmented image.
image, as shown in Fig. 5. The comparison among the ground
truth image and the proposed segmented image is defined DCNN features and recognition
through Eq. (7).
In the area of machine learning (ML), the performance of
a system is dependent on the extraction of most strong
jξGR ðiÞ∩ξRG ð jÞj
ξbd ði; jÞ ¼ 1− ð7Þ kind features [41–45]. In medical imaging, mostly color,
jξGR ðiÞj þ jξRG ð jÞj−jξGR ðiÞ∩ξRG ð jÞj texture, and shape features are used for the classification

Table 1 Segmentation accuracy

for ISIC 2017 dataset using Image No Accuracy (%) Negative rate (%) Image No Accuracy (%) Negative rate (%)
testing data; (* denotes the overall
segmentation accuracy of all 1 98.25 1.75 16 88.47 11.53
training images) 2 96.56 3.44 17 88.91 11.09
3 96.12 3.88 18 89.17 10.83
4 94.27 5.73 19 89.92 10.08
5 95.95 4.05 20 90.62 9.38
6 95.89 4.11 21 92.91 7.09
7 95.14 4.86 22 92.98 7.02
8 97.94 2.06 23 91.67 8.33
9 97.61 2.39 24 91.88 8.12
10 97.29 2.71 25 93.79 6.21
11 96.80 3.2 26 92.35 7.65
12 96.42 3.58 27 91.51 8.49
13 96.91 3.09 28 94.34 5.66
14 96.97 3.03 29 94.66 5.34
15 96.56 3.44 30 94.80 5.2
*Overall Average 94.78 5.22
289 Page 10 of 19 J Med Syst (2019) 43: 289

Table 2 Segmentation accuracy

for PH2 dataset using testing data; Image No Accuracy (%) Negative rate (%) Image No Accuracy (%) Negative rate (%)
(* denotes the overall
segmentation accuracy of all 1 98.11 1.89 11 94.01 5.99
training images) 2 98.19 1.81 12 94.19 5.81
3 98.71 1.29 13 94.89 5.11
4 98.75 1.25 14 94.99 5.01
5 98.80 1.2 15 97.56 2.44
6 98.88 1.12 16 97.11 2.89
7 96.59 3.41 17 97.89 2.11
8 96.41 3.59 18 97.90 2.1
9 96.11 3.89 19 97.95 2.05
10 94.12 5.88 20 97.98 2.02
*Overall Average 95.41 4.59

process [46–50], but these techniques do not perform well testing. In the training phase, the input dataset is divided into
when the processed data is vast in number. Researchers a 70:30 which explains that 70% of samples are utilized for
resolve this kind of problem through deep learning in the training the model and the remaining 30% for testing the pro-
area of ML [46, 51–53]. In this work, we utilized a pre- posed system. The dimensions of the features extracted
trained inception V3 DCNN model based on its high through FC and AP layers are Nx1000 and Nx20148 respec-
speed and best accuracy. The original architecture of tively. Later, the extracted features from both layers are fused
Inception V3 is shown in Fig. 6 [54]. through the Hamming Distance (HD) approach as described
Inception V3 CNN pre-trained model is introduced after below.
the modification of GoogleNet architecture. The base of Let X denotes the number of the extracted features repre-
this model is replaced by 3 × 3 convolutions whereas the sented as Fcland APl. Let the values of the extracted features
previous GoogleNet has 5 × 5 convolutions. The resultant are Δ = {x1, x2, …xn − 1}. Then Δ is representable according
network includes ten inception modules. The modules of to a bijection f : {ci, cj} ↔ Δ. We defined the distortion D(f,
GoogleNet are changed by substituting the nxn convolu- ci, cj) under the bijectionf between two feature sets ci, cj ∈ {R}n
tions by 1 × 7 followed by a 1 × 1 which reduced the com- as the normalized D1,the distance of their corresponding fea-
putational time. The last two modules of GoogleNet are tures.
replaced by 1 × 3 and 3 × 1 in parallel. Lastly, the original   
7 × 7 convolutions of GoogleNet model are modified by    f ðci Þ−f c j 
D f ; ci ; c j ¼ ð8Þ
3 × 3 convolutions. Hence, in total 42 learnable layers are 2X
included in Inception v3 CNN model. Where, ci, cj ∈ {R}n and R represent the real values of the
Features are extracted through performing the activation on extracted features. Let DH(xi, xj) be the HD between two fea-
FC and AP layers of Inception V3 model in parallel through ture sets ciand cj, then the maximum distance between two
fine-tuning. This process includes two-step as training and feature sets is assumed as follows:

Table 3 Comparison of proposed segmentation accuracy for ISIC 2017 DH ðmax;1Þ ð f ; cÞ ¼ max Dð f ; c; xÞ ð9Þ
fc:DH ðc;xÞ¼1g
and PH2 dataset

Method Dataset Year Accuracy (%) Then, the distance-1 distortion is defined under bijection f
as:
[56] ISIC 2017 2019 94.08
[59] ISIC 2017 2018 85.40 DH ðmax;1Þ ð f Þ ¼ maxn DH ðmax;1Þ ð f ; cÞ ð10Þ
c∈fRg
[60] ISIC 2017 2018 83.90
[61] ISIC 2017 2019 93.80
[62] ISIC 2017 2017 91.00 Where, DH(max, 1) defines the max distance features that re-
[56] PH2 2019 95.30 places the lower distance value feature at fusion place. The
[63] PH2 2017 94.24 size of the fused vector is N × 2048 denoted by δF which later
Proposed ISIC 2017 2019 94.78 is selected through anew implemented clustering controlled
Proposed PH2 2019 95.41
entropy (CcE) approach. The visual representation of the fu-
sion process is shown in Fig. 7.
J Med Syst (2019) 43: 289 Page 11 of 19 289

Table 4 Proposed recognition results through HD based features fusion and CcE selection approach using PH2 dataset

Classifier Method Measures

CT ST Sen (%) Spec (%) Prec (%) F1 M (%) AUC Acc (%) FPR Time (sec)

Trees ✓ 77 75 77.3 77.15 0.85 77.1 0.117 16.80

✓ 78.75 76.5 78.0 78.37 0.86 78.1 0.11 4.29
DA LDA QDA Measures
✓ 98.15 98.3 98.1 98.13 0.98 98.2 0.01 15.64
✓ 91.67 98.5 92.7 92.18 0.96 91.7 0.04 4.04
SVM LSVM CSVM QSVM MGSVM Measures
✓ 97.67 97.5 97.7 97.68 0.99 97.9 0.01 7.29
✓ 91.70 98.5 92.7 92.2 0.97 91.8 0.05 9.16
✓ 97.3 98.5 97.3 97.3 0.99 97.4 0.02 5.30
✓ 90.67 92.0 90.7 90.68 0.96 90.6 0.467 5.21
KNN FKNN MKNN Cosine Cubic WKNN Measures
✓ 96.5 99 97.0 96.75 0.98 97.4 0.013 4.882
✓ 94.75 98.5 96.0 95.37 0.99 95.8 0.023 4.04
✓ 94.0 97.0 95.0 94.49 0.99 94.8 0.267 4.19
✓ 93.5 99 95.7 94.59 0.99 95.3 0.023 5.97
✓ 97.4 99.0 97.3 97.35 0.99 97.4 0.02 4.02
NN MLP 98.25 98.5 98.3 98.27 1.0 98.4 0.01 3.25
EBT Baggage tree 90.67 92.0 91.3 90.98 0.97 91.1 0.043 10.03

*The bold results of Neural Network show the best accuracy

The fusion process includes few drawbacks such as the the length of fused features. We used K-Means clustering on
addition of irrelevant features, redundancy among fea- fused FV with k = 2. As, we defined the clusters {Ck, k = 1, 2}.
tures, and few more. These kinds of problems make the Then, the posterior probability of the defined clusters is com-
recognition process complex and time-consuming [55]. puted as follows:
Few evolutionary algorithms are reported in state of art
PðC k jδ F Þ ¼ PðδFjC k ÞPðC k ÞjPðδF Þ∝PðδFjC k ÞPðC k Þ ð11Þ
using GA and PSO. However, these techniques are time-
consuming to optimize fitness function. Compared to The cost function of K-Means clustering is defined as fol-
these techniques, the physics approaches like entropy lows:
methods are fast and efficient on a large number of h i
datasets. In this work, we implemented a CcE approach ∑λi ∈Ck ðψi −C k Þ2 =2σ2 þ ðδF Þ ð12Þ
to select the best features. Mathematically, the CcE tech-
nique is represented in Eq. 11. After that, we computed the entropy of each cluster and
Let we have fused feature vector (FV) denoted by δF of selected the highest value features through weights posterior
dimension N × 2048, where, N denotes the number of samples probability assignment. The entropy function of this work is
that are used for training and testing process and 2048 denotes modified as:
   
PðδFjC k Þ ¼ exp W∑λi ∈Ck ln P ψi jC k ð13Þ
Table 5 Verification for classification results of the neural network in
the form of CM Where, ψi denotes the expected outcome features, W is a
Cancer type Cancer type weight parameter which is formulated as W ¼ Cnk .
Finally, top features are selected according to the values
Common Melanoma Benign
that are higher than W. The selected features are finally clas-
Common 100% sified through neural network (NN) [4]. The performance of
Melanoma 98% 2% these results is compared with several other classification al-
Benign 3% 97% gorithms like SVM, KNN, and few more [10–12]. The sample
results of the proposed workflow are shown in Fig. 8.
289 Page 12 of 19 J Med Syst (2019) 43: 289

Fig. 10 Presentation of PH2 dataset results in the form of ROC curves

Experimental results and discussion different types of classifiers such as decision trees (DT), dis-
criminant analysis (LDA), support vector machine (SVM), k-
Experimental Setup nearest neighbor (KNN), ensemble methods, and neural net-
work (NN). We utilized neural network in this work for its best
The proposed skin lesion identification and recognition sys- results and compared its performance with other state-of-the-
tem are evaluated on three publically dermoscopic datasets art classifiers. The classifiers mentioned above are utilized in
including PH2, and ISBI series (2016 and 2017). The exper- the classification phase. They also include a few advanced
imental results are obtained in two steps. In the first step, techniques like simple trees, fine trees, cubic SVM, weighted
lesion identification is performed through the proposed KNN, ensemble subspace discriminant analysis (ESDA), and
contrast-enhanced CNN along with the thresholding. The ex- few more as defined in the results section. The analysis of NN
periments are achieved on PH2 and ISIC 2017 datasets. In the is performed on the selected datasets through 8 performance
second phase, the best-selected features are classified through metrics like sensitivity (sen), specificity (spec), precision

Table 6 Proposed recognition results through HD based features fusion and CcE selection approach using ISBI 2016 dataset

Classifier Method Measures

CT ST Sen Spec Prec F1 M AUC Acc FPR Time

(%) (%) (%) (%) (%) (sec)

Trees ✓ 90.5 90 90.5 90.5 0.93 90.8 0.095 13.68

✓ 91.5 94 92.0 91.69 0.91 91.8 0.085 9.34
Discriminant LDA QDA Measures
✓ 88.5 83 89 88.75 0.92 88.5 0.115 9.58
✓ 88.0 91.5 91.5 89.72 0.94 91.4 0.085 8.39
Support Vector LSVM CSVM QSVM MGSVM CGSVM Measures
Machine ✓ 94.0 91.0 94.0 94 0.98 93.8 0.06 9.84
✓ 92.5 93.0 92.5 92.5 0.97 92.4 0.15 8.70
✓ 91.5 92.0 92.0 91.75 0.98 91.8 0.08 8.75
✓ 92.0 90.0 92.5 92.25 0.97 92.1 0.08 8.50
✓ 89.0 83.0 89.5 89.25 0.96 88.8 0.11 8.37
K-Nearest Neighbor FKNN MKNN Cosine Cubic W-KNN Measures
✓ 86.0 90.0 90.5 88.20 0.90 89.8 0.10 8.59
✓ 87.5 78.0 89.5 88.5 0.96 87.8 0.125 8.03
✓ 86.0 90.5 91.0 88.42 0.97 90.8 0.095 8.59
✓ 89.5 83.0 90.0 89.75 0.96 89.5 0.105 10.67
✓ 89.0 81.0 90.5 89.75 0.96 89.1 0.105 7.97
Neural Network MLP 95.0 95.0 95.0 95.0 0.98 95.1 0.05 6.97
EBT Baggage tree 92.0 91.0 92.5 92.25 0.97 92.1 0.08 20.59

*The bold results of Neural Network show the best accuracy

J Med Syst (2019) 43: 289 Page 13 of 19 289

Table 7 Verification for proposed system [58]. The few of sample images are
classification results of Cancer type Cancer type
shown in Fig. 9.
the neural network in the
form of CM Benign Malignant

Benign 95% 5% Segmentation results

Malignant 5% 95%
The two well-known performance metrics: accuracy and neg-
ative rate (NR) are calculated for performance analysis of
(pre), F1-measure (F1-M), AUC, false positive rate (FPR), lesion segmentation. The ISIC 2017 and PH2 datasets are
accuracy (Acc), and recognition execution time. All experi- utilized in this regards. The annotated images of both datasets
ments are performed on MATLAB using Desktop Computer are provided by experts and are publically available. The total
Corei7 with 16GB of RAM and 8GB graphic card. 2000 images of ISIC 2017 dataset is utilized, and a few best
results (30 images) are given in Table 1. The overall average
Datasets accuracy achieved for all 2000 images is 94.98%. The overall
error rate is 5.22 which is less (almost 1%) as compared to the
In this work, we utilized three famous dermoscopy (DY) latest state of the art techniques.
datasets; ISBI 2016, ISBI 2017, and PH2. The PH2 dataset Similarly, the segmentation results of PH2 dataset are also
given in Table 2. The total of 200 DY samples experiments;
consists of a total of 200 8-bit DY images of resolution 768 ×
660. The total of 200 DY images includes 80 benign, 80 com- however, results are exhibited on few images. The overall average
mon nevi, and 40 melanoma. In addition, the ground truth accuracy of all 200 images is 95.41% and the error rate is 4.59%
which is less (almost 0.52%) as compared to the latest techniques.
images of all 200 DY images are also provided by expert
doctors and publically available for the segmentation process The comparison of the proposed segmentation results with
[56]. the latest techniques is conducted in Table 3. The recent max-
imum achieved accuracy of ISIC 2017 dataset is 94.08% [56].
The ISBI 2016 challenge includes a total of 1279 DY im-
ages. This challenge includes separate tasks such as segmen- Whereas, our method shows improvement in accuracy up to
tation, features detection, and classification of lesions into the 0.7% and reached 94.78% for ISIC 2017 dataset. Similarly,
for PH2 dataset, the previous best-reported accuracy is
relevant category. For the training process, 900 images are
utilized from total DY samples and 379 images for the testing 95.30% [56] and our method reached 95.41% accuracy which
process [57]. shows 0.11% improvement.
The ISBI 2017 challenge includes the total of 2750 DY
RGB images of resolution 296-1456. From the total, 2233 Recognition results
images are benign, and 517 are melanoma. For the exper-
imental evaluation, 2000 images are utilized for the train- In the lesion recognition task, three datasets are utilized for the
ing process and the remaining 750 for testing the system analysis. The extracted CNN features are fused and

Fig. 11 Presentation of ISBI 2016 dataset results in the form of ROC curves
289 Page 14 of 19 J Med Syst (2019) 43: 289

Table 8 Proposed recognition results through HD based features fusion and CcE selection approach using ISBI 2017 dataset

*Meth Method Measures

CT ST Sen (%) Spec (%) Prec (%) F1 M (%) AUC Acc (%) FPR Time (sec)

Trees ✓ 91.5 93 91.5 91.5 0.95 91.7 0.085 8.87

✓ 88.5 85 89 88.75 0.91 88.7 0.115 12.15
DA LDA QDA Measures
✓ 94 95 94 94 0.94 94.3 0.06 12.4
✓ 93.3 99.1 93.1 93.45 0.96 93.4 0.08 24.96
SVM L C Q-SVM MG-SVM CG-SVM Measures
SVM SVM
✓ 94 99 94.5 94.25 0.97 94.1 0.06 10.9
✓ 85.5 87 85.5 85.5 0.92 85.6 0.14 139.43
✓ 94 99 94.5 94.25 0.98 94.0 0.06 22.52
✓ 94 98 94 94 0.98 94.0 0.06 15.64
✓ 93 99 93.5 93.25 0.97 93.0 0.07 14.59
KNN FKNN M-KNN Cosine Cubic W KNN Measures
✓ 92.5 94 92.5 92.5 0.92 92.5 0.07 10.68
✓ 94 100 94.5 94.25 0.97 93.8 0.06 9.898
✓ 94 99 94.5 94.25 0.98 94.1 0.05 10.89
✓ 93 99 93.5 93.25 0.97 93.1 0.07 99.94
✓ 99 100 94.5 96.7 0.97 94.1 0.06 14.81
NN MLP 94.5 98 95 94.75 0.98 94.8 0.05 10.91
EBT Baggage tree 94 99 94.5 94.25 0.98 94.0 0.06 22.52

*The bold results of Neural Network show the best accuracy. *Meth denotes method or classifiers

best subset of features is selected through CcE based ap- FPR are 98.25%, 98.5%, 8.3%, 98.27%, 1.0%, and 0.01,
proach. Later, the selected best features are classified using respectively. The best accuracy values of the other classi-
NN and few other state-of-the-art algorithms for comparison fiers such as DT, DA, SVM, KNN, and EBT are 78.1%,
purpose. In addition, the advanced methods of these classifiers 98.2% 97.9%, 97.4%, and 91.1% using ST, LDA, LSVM,
such as complex tree and simple tree are used for DT. W-KNN, and baggage trees, respectively. The recognition
Similarly, for discriminant analysis, the LDA and QDA are accuracy of NN is also verified through a confusion ma-
employed, for SVM, LSVM, CSVM, QSVM, and MGSVM trix (CM) and ROC curves as presented in Table 5 and
are implemented. A 70:30 approach is utilized to divide the Fig. 10. In addition, the execution time of the recognition
datasets into training and testing procedures. Then, k-fold process is computed for all classifiers and the top three
cross validation is performed where k = 20 which normally best recorded times are 3.25 s, 4.02 s, and 4.04 s for NN,
selected as k = 10, but we selected this value for more reliable weighted KNN, medium KNN, respectively.
and consistent results. The recognition accuracy values of ISBI 2016 dataset
The recognition performance of PH2 dataset in terms for all the selected classifiers are presented in Table 6.
of a different number of classifiers and their performance The best-noted accuracy is 95.1% using NN through
matrices are presented in Table 4. In this Table, the best- MLP method. The other achieved measures like Sen,
achieved accuracy is 98.4% for NN, whereas the other Spec, Prec, F1-M, AUC, and FPR are 95.0%, 95.0%,
parameters such as Sen, Spec, Pre, F1-M, AUC, and 95.0%, 95%, 0.98, and 0.05, respectively. The other clas-
sifiers such as DT, DA, SVM, KNN, and EBT achieved
Table 9 Verification for the best accuracy of 91.8%, 91.4%, 93.8%, 90.8%, and
classification results of Cancer type Cancer type 92.1% using ST, QDA, LSVM, cosine-KNN, and baggage
the neural network in the trees, respectively. The recognition accuracy of NN using
form of CM Benign Malignant
MLP method is also verified through a confusion matrix
Benign 98% 2% (CM) and ROC curves as presented in Table 7 and
Malignant 9% 91% Fig. 11. In addition, the execution time of the recognition
process is also computed for all the selected classification
J Med Syst (2019) 43: 289 Page 15 of 19 289

Fig. 12 Presentation of ISBI 2017 dataset results in the form of ROC curves

methods, and the best recorded time is 6.97 s. The other used for verification of the proposed accuracy using NN. In
two best execution performances are 7.97 and 8.03 s for addition, the confirmation of NN achievement is also shown
W-KNN and MKNN methods, respectively. through ROC curves in Fig. 12. The comparison of the run-
In Table 8, the detailed recognition results are presented for ning time is also carried out for all the selected classifiers in
ISBI 2017 dataset. Similar to the PH2 and ISBI 2016 datasets Table 9 using the proposed method. The best reported com-
evaluation process, the same classifiers and performance mea- putational time is 8.876 for decision trees (DT) method like a
sures are employed for this dataset. The best-recorded accura- complex tree (CT).
cy is 94.8% which is obtained through NN. The other calcu-
lated measures like Sen, Spec, Prec, F1-M, AUC, and FPR are Comparison and Analysis
94.5%, 98%, 95%, 94.75%, 0.98, and 0.055%, respectively.
The best accuracy values of DT, DA, SVM, KNN, and EBT In this section, the analysis of the proposed approach is con-
are 91.7%, 94.3%, 94.1%, 94.1%, and 94%, respectively ducted in terms of both numerical values and visual results
using CT, LDA, LSVM, weighted KNN, and baggage tree like enhancement, lesion boundary extraction, and labeled im-
methods, respectively. In Table 9, the CM is given which is ages. The overall architecture of the proposed system is shown

Fig. 13 Graphical comparison of

best accuracies of each
classification method for all three
selected datasets
289 Page 16 of 19 J Med Syst (2019) 43: 289

Fig. 14 Graphical representation

of classification time which are
listed in Fig. 13 for all datasets

in Fig. 1 whereas the visual enhancement and DCNN based The proposed method is also compared with the latest tech-
segmentation effects are shown in Figs. 2, 3, 4, and 5. The niques using the three selected datasets in term of accuracy as
proposed segmentation method is validated on ISIC 2017 and addressed in Table 10. In [4], the authors presented the
PH2 datasets and the obtained results are presented in Tables 1 handcrafted features and the entropy-based selection approach
and 2. In the recognition phase, the CNN features are extracted for lesion classification. They reached an accuracy of 97.5%
and fused through HD based approach and the visual effects using PH2 dataset. Waheed et al. [64] presented an efficient
are shown in Fig. 7. The overall labeled recognition results are machine learning approach for skin lesion classification and
shown in Fig. 8. The recognition process is validated on ISBI reported an accuracy of 96% using PH2 dataset. Similarly in
2016, ISBI 2017, and PH2 datasets and the obtained results [26, 69], the authors introduced the hand-crafted and the deep
are presented in Tables 3-9 and their ROC’s are displayed in CNN based methods for lesion classification and achieved an
Figs. 10-12. In Fig. 13, a comparison of each classification accuracy of 97.5% and 97%, respectively. The proposed
method for all three selected datasets is presented which method achieves an accuracy of 98.4% on PH2 dataset. Like
shows the superiority of NN compared to the other techniques. PH2 dataset, the comparison is also conducted for ISBI 2016
In addition, the classification time is also plotted in Fig. 14. and 2017 datasets with the latest approaches ([68]; M Attique
From Fig. 14, NN performed efficiently compared to the other [4, 8, 66, 67]). Recently, Yu et al. [8] achieved an accuracy of
methods. 94.9% using ISBI 2016 dataset, whereas our method achieves
an accuracy of 95.1%. In [68], Bi et al. reported an accuracy of
93.4% which is later improved by [67] with more than 0.2%.
Table 10 Comparison with existing techniques for all selected datasets The proposed method achieves an accuracy of 94.8% for ISBI
Method Year Dataset Accuracy (%)
2017 dataset which gives the validity of the proposed deep
learning method.
[4] 2018 PH2 97.5
[64] 2017 PH2 96.0
[65] 2018 PH2 97.5 Conclusion
[26] 2017 PH2 97.0
Proposed 2019 PH2 98.4 In this article, we proposed a new fully automated enhanced
[4] 2018 ISBI 2016 83.2 deep CNN model for skin lesion border identification and
[66] 2018 ISBI 2016 86.1 lesion recognition. The proposed heterogeneous framework
[8] 2017 ISBI 2016 94.9 includes three major pipeline procedures including augmenta-
Proposed 2019 ISBI 2016 95.1 tion and contrast enhancement, color pixels based on CNN
[4] 2018 ISBI 2017 88.2 training and segmentation, and CNN features fusion and se-
[67] 2018 ISBI 2017 93.6 lection. The best subset of selected features is classified
[68] 2017 ISBI 2017 93.4 through NN. A comparison result is conducted for each
Proposed 2019 ISBI 2017 94.8 dataset using several classification methods. The segmenta-
tion results are evaluated on PH2 and ISIC 2017 datasets
J Med Syst (2019) 43: 289 Page 17 of 19 289

and attained an average accuracy of 95.41% and 94.78%, 8. Yu, L., Chen, H., Dou, Q., Qin, J., and Heng, P.-A., Automated
melanoma recognition in dermoscopy images via very deep resid-
respectively. The proposed recognition process is evaluated
ual networks. IEEE Transactions on Medical Imaging 36(4):994–
on ISIB 2016, 17, and PH2 datasets and attained the best 1004, 2017.
accuracy of 95.1%, 94.8%, and 98.4%, respectively. The pro- 9. Korotkov, K., and Garcia, R., Computerized analysis of pigmented
posed system accuracy is better as compared to the existing skin lesions: a review. Artificial Intelligence in Medicine 56(2):69–
state-of-the-art segmentation and recognition techniques 90, 2012.
10. Khan, M. A., Akram, T., Sharif, M., Saba, T., Javed, K., Lali, I. U.
which shows the authenticity of our method. Moreover, from et al., Construction of saliency map and hybrid set of features for
the obtained results, we conclude that the selection of the most efficient segmentation and classification of skin lesion. Microsc Res
relevant pixels of lesion regions provides good segmentation Tech, 2019a. https://doi.org/10.1002/jemt.23220.
results that later affect the recognition process. We also notice 11. Khan, S. A., Nazir, M., Khan, M. A., Saba, T., Javed, K., Rehman,
that the selection of the best subset of features reduces the A., Akram, T., and Awais, M., Lungs nodule detection framework
from computed tomography images using support vector machine.
overall system execution time. In the future, we shall attain Microsc Res Tech., 2019b. https://doi.org/10.1002/jemt.23275.
only the segmentation process using transferring learning and 12. Khan, M. A., Lali, I. U., Rehman, A., Ishaq, M., Sharif, M., Saba,
physics selection theorems. Through physics selection theo- T., Zahoor, S., and Akram, T., Brain tumor detection and classifi-
rems, the best subsets of features are selected. cation: A framework of marker-based watershed algorithm and
multilevel priority features selection. Microscopy Research and
Technique, 2019c. https://doi.org/10.1002/jemt.23238.
Acknowledgements This work was supported by the research Project
13. Yousaf, K., Mehmood, Z., Saba, T., Rehman, A., Munshi, A. M.,
[Skin Cancer Melanoma Detection from Dermoscopic Images Using
Alharbey, R., and Rashid, M., Mobile-health applications for the
Machine Learning Techniques]; Prince Sultan University; Saudi Arabia
efficient delivery of health care facility to people with dementia
[SSP -18-5-04]Additionally, in part supported by Artificial Intelligence
(PwD) and support to their carers: A survey. BioMed Research
and Data Analytics (AIDA) Lab Prince Sultan University Riyadh Saudi
International 2019:1–26, 2019.
Arabia. Authors are thankful for the support.
This work was supported by the Research Project (SSP -18-5-04). 14. Suzuki, K., Overview of deep learning in medical imaging.
Additionally, in part supported by Artificial Intelligence and Data Radiological Physics and Technology 10(3):257–273, 2017.
Analytics (AIDA) Lab Prince Sultan University Riyadh Saudi Arabia. 15. Shen, D., Wu, G., and Suk, H.-I., Deep learning in medical image
Authors are thankful for the support. analysis. Annual Review of Biomedical Engineering 19:221–248,
2017.
16. Ching, T., Himmelstein, D. S., Beaulieu-Jones, B. K., Kalinin, A.
A., Do, B. T., Way, G. P. et al., Opportunities and obstacles for deep
References learning in biology and medicine. Journal of The Royal Society
Interface 15(141):20170387, 2018.
17. Norouzi, A., Rahim, M. S. M., Altameem, A., Saba, T., Rada, A. E.,
1. Rogers, H. W., Weinstock, M. A., Feldman, S. R., and Coldiron, B.
Rehman, A., and Uddin, M., Medical image segmentation methods,
M., Incidence estimate of nonmelanoma skin cancer (keratinocyte
algorithms, and applications. IETE Technical Review 31(3):199–
carcinomas) in the US population, 2012. JAMA Dermatology
213, 2014. https://doi.org/10.1080/02564602.2014.906861.
151(10):1081–1086, 2015.
18. Lee, T., Ng, V., Gallagher, R., Coldman, A., and McLean, D.,
2. Jerant, A. F., Johnson, J. T., Demastes Sheridan, C., and Caffrey, T.
Dullrazor®: A software approach to hair removal from images.
J., Early detection and treatment of skin cancer. American family
Computers in Biology and Medicine 27(6):533–543, 1997.
physician 62(2), 2000.
19. Kiani, K., and Sharafat, A. R., E-shaver: An improved DullRazor®
3. Balch, C. M., Gershenwald, J. E., Soong, S.-J., Thompson, J. F.,
for digitally removing dark and light-colored hairs in dermoscopic
Atkins, M. B., Byrd, D. R. et al., Final version of 2009 AJCC
images. Computers in Biology and Medicine 41(3):139–145, 2011.
melanoma staging and classification. Journal of clinical oncology
20. Liu, Z., and Zerubia, J., Skin image illumination modeling and
27(36):6199, 2009.
chromophore identification for melanoma diagnosis. Physics in
4. Khan, M. A., Akram, T., Sharif, M., Awais, M., Javed, K., Ali, H.,
Medicine & Biology 60(9):3415, 2015.
and Saba, T., CCDF: Automatic system for segmentation and rec-
21. Emre Celebi, M., Kingravi, H. A., Iyatomi, H., Alp Aslandogan, Y.,
ognition of fruit crops diseases based on correlation coefficient and
Stoecker, W. V., Moss, R. H. et al., Border detection in dermoscopy
deep CNN features. Computers and Electronics in Agriculture 155:
images using statistical region merging. Skin Research and
220–236, 2018.
Technology 14(3):347–353, 2008.
5. Nasir, M., Khan, M. A., Sharif, M., Lali, I. U., Saba, T., Iqbal, T.,
22. Gómez, D. D., Butakoff, C., Ersboll, B. K., and Stoecker, W.,
An improved strategy for skin lesion detection and classification
Independent histogram pursuit for segmentation of skin lesions.
using uniform segmentation and feature selection based approach,
IEEE Transactions on Biomedical Engineering 55(1):157–161,
Microscopy research and technique 81(6):528–543, 2018. https://
2008.
doi.org/10.1002/jemt.23009
23. Silveira, M., Nascimento, J. C., Marques, J. S., Marçal, A. R.,
6. Binder, M., Schwarz, M., Winkler, A., Steiner, A., Kaider, A.,
Mendonça, T., Yamauchi, S. et al., Comparison of segmentation
Wolff, K., and Pehamberger, H., Epiluminescence microscopy: a
methods for melanoma diagnosis in dermoscopy images. IEEE
useful tool for the diagnosis of pigmented skin lesions for formally
Journal of Selected Topics in Signal Processing 3(1):35–45, 2009.
trained dermatologists. Archives of Dermatology 131(3):286–291,
1995. 24. Yao, Q., Guan, Z., Zhou, Y., Tang, J., Hu, Y., & Yang, B. (2009).
Application of support vector machine for detecting rice diseases
7. Abbas, Q., Garcia, I., and Rashid, M., Automatic skin tumor border
using shape and color texture features. Paper presented at the
detection for digital dermoscopy using a new digital image analysis
Engineering Computation, 2009. ICEC'09. International
scheme. British Journal of Biomedical Science 67(4):177–183,
Conference on.
2010.
289 Page 18 of 19 J Med Syst (2019) 43: 289

25. Sadeghi, M., Lee, T. K., McLean, D. I., Lui, H., and Atkins, M. S., due to breast tumor. Microscopy Research and Technique, 2017b.
Detection and analysis of irregular streaks in dermoscopic images https://doi.org/10.1002/jemt.22961.
of skin lesions. IEEE Trans. Med. Imaging 32(5):849–861, 2013. 44. Saba, T., Khan, S. U., Islam, N., Abbas, N., Rehman, A. et al.,
26. Satheesha, T., Satyanarayana, D., Prasad, M. G., and Dhruve, K. D., Cloud-based decision support system for the detection and classifi-
Melanoma is Skin Deep: A 3D reconstruction technique for com- cation of malignant cells in breast cancer using breast cytology
puterized dermoscopic skin lesion classification. IEEE Journal of images. Microscopy Research and Technique, 2019. https://doi.
Translational Engineering in Health and Medicine 5:1–17, 2017. org/10.1002/jemt.23222.
27. Celebi, M. E., and Zornberg, A., Automated quantification of clin- 45. Saba, T., Al-Zahrani, S., and Rehman, A., Expert system for offline
ically significant colors in dermoscopy images and its application to clinical guidelines and treatment. Life Sci Journal 9(4):2639–2658,
skin lesion classification. IEEE systems journal 8(3):980–984, 2012.
2014. 46. Iqbal, S., Khan, M. U. G., Saba, T., Mehmood, Z., Javaid, N.,
28. Tschandl, P., Sinz, C., and Kittler, H., Domain-specific classifica- Rehman, A., and Abbasi, R., Deep learning model integrating fea-
tion-pretrained fully convolutional network encoders for skin lesion tures and novel classifiers fusion for brain tumor segmentation.
segmentation. Computers in biology and medicine 104:111–116, Microscopy Research and Technique, 2019. https://doi.org/10.
2019. 1002/jemt.23281.
29. Mahbod, A., Schaefer, G., Ellinger, I., Ecker, R., Pitiot, A., and 47. Saba, T., Rehman, A., Mehmood, Z., Kolivand, H., and Sharif,
Wang, C., Fusing fine-tuned deep features for skin lesion classifi- M., Image Enhancement and Segmentation Techniques for
cation. Computerized Medical Imaging and Graphics 71:19–29, Detection of Knee Joint Diseases: A Survey. Current Medical
2019. Imaging Reviews 14(5), 2018b. https://doi.org/10.2174/
30. Rebouças Filho, P. P., Peixoto, S. A., da Nóbrega, R. V. M., 1573405613666170912164546.
Hemanth, D. J., Medeiros, A. G., Sangaiah, A. K., and de 48. Saba, T., Bokhari, S. T. F., Sharif, M., Yasmin, M., and Raza, M.,
Albuquerque, V. H. C., Automatic histologically-closer classifica- Fundus image classification methods for the detection of glaucoma:
tion of skin lesions. Computerized Medical Imaging and Graphics., A review. Microsc Res Tech., 2018a. https://doi.org/10.1002/jemt.
2018. 23094.
31. Oliveira, R. B., Papa, J. P., Pereira, A. S., and Tavares, J. M. R., 49. Sadad, T., Munir, A., Saba, T., and Hussain, A., Fuzzy C-means and
Computational methods for pigmented skin lesion classification in region growing based classification of tumor from mammograms
images: review and future trends. Neural Computing and using hybrid texture feature. Journal of Computational Science 29:
Applications 29(3):613–636, 2018. 34–45, 2018.
32. Ray, S. (2018). Disease Classification within Dermascopic Images 50. Ullah, H., Saba, T., Islam, N., Abbas, N., Rehman, A., Mehmood,
Using features extracted by ResNet50 and classification through Z., Anjum, A., An ensemble classification of exudates in color
Deep Forest. arXiv preprint arXiv:1807.05711. fundus images using an evolutionary algorithm based optimal fea-
tures selection, Microscopy research and technique, 2019. https://
33. Eaton-Rosen, Z., Bragman, F., Ourselin, S., & Cardoso, M. J.
doi.org/10.1002/jemt.23178.
(2018). Improving Data Augmentation for Medical Image
51. Iqbal, S., Ghani, M. U., Saba, T., and Rehman, A., Brain tumor
Segmentation.
segmentation in multi-spectral MRI using convolutional neural net-
34. Paris, S., Hasinoff, S. W., and Kautz, J., Local Laplacian filters:
works (CNN). Microsc Res Tech. 81(4):419–427, 2018. https://doi.
Edge-aware image processing with a Laplacian pyramid. ACM
org/10.1002/jemt.22994.
Trans. Graph. 30(4):61–68, 2011 12.
52. Iqbal, S., Khan, M. U. G., Saba, T., and Rehman, A., Computer
35. Kaur, R., Stoecker, W. V. D., Mishra, N. K., & Kasmi, R. (2018).
assisted brain tumor type discrimination using magnetic resonance
Thresholding methods for lesion segmentation in dermoscopy im-
imaging features. Biomedical Engineering Letters 8(1):5–28, 2017.
ages: Google Patents.
https://doi.org/10.1007/s13534-017-0050-3.
36. Guo, Y., Ashour, A. S., and Smarandache, F., A Novel Skin Lesion 53. Tahir, B. Iqbal, S., Khan, M.U.G., Saba, T., Mehmood, Z., Anjum,
Detection Approach Using Neutrosophic Clustering and Adaptive A., Mahmood, T. (2019) Feature enhancement framework for brain
Region Growing in Dermoscopy Images. Symmetry 10(4):119, tumor segmentation and classification, https://doi.org/10.1002/
2018. jemt.23224.
37. Xu, H., & Hwang, T. H. (2018). Automatic Skin Lesion 54. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., & Wojna, Z.
Segmentation Using Deep Fully Convolutional Networks. arXiv (2016). Rethinking the inception architecture for computer vision.
preprint arXiv:1807.06466. Paper presented at the Proceedings of the IEEE conference on com-
38. Jamal, A., Hazim Alkawaz, M., Rehman, A., and Saba, T., (2017) puter vision and pattern recognition.
Retinal imaging analysis based on vessel detection. Microsc Res 55. Sharif, M., Khan, M. A., Akram, T., Javed, M. Y., Saba, T., and
Tech. 80(17):799–811, 2017. https://doi.org/10.1002/jemt. Rehman, A., A framework of human detection and action recogni-
39. Rahim, M. S. M., Rehman, A., Kurniawan, F., and Saba, T., Ear tion based on uniform segmentation and combination of Euclidean
biometrics for human classification based on region features min- distance and joint entropy-based features selection. EURASIP
ing. Biomedical Research, vol. 28(10):4660–4664, 2017. Journal on Image and Video Processing 2017(1):89, 2017.
40. Duan, Q., Akram, T., Duan, P., & Wang, X., Visual saliency detec- 56. Bi, L., Kim, J., Ahn, E., Kumar, A., Feng, D., and Fulham, M.,
tion using information contents weighting. Optik 127(19):7418– Step-wise integration of deep class-specific learning for
7430, 2016. dermoscopic image segmentation. Pattern recognition 85:78–89,
41. Mughal, B., Muhammad, N., Sharif, M., Rehman, A., and Saba, T., 2019.
Removal of pectoral muscle based on topographic map and shape- 57. Wu, J. T., Dernoncourt, F., Gehrmann, S., Tyler, P. D., Moseley, E.
shifting silhouette. BMC Cancer, 2018. https://doi.org/10.1186/ T., Carlson, E. T. et al., Behind the scenes: A medical natural lan-
s12885-018-4638-5. guage processing project. International Journal of Medical
42. Mughal, B., Muhammad, N., Sharif, M., Saba, T., and Rehman, A., Informatics 112:68–73, 2018.
Extraction of breast border and removal of pectoral muscle in wave- 58. Codella, N. C., Gutman, D., Celebi, M. E., Helba, B., Marchetti, M.
let, domain. Biomedical Research 28(11):5041–5043, 2017a. A., Dusza, S. W., … Kittler, H. (2018). Skin lesion analysis toward
43. Mughal, B., Sharif, M., Muhammad, N., and Saba, T., A novel melanoma detection: A challenge at the 2017 international sympo-
classification scheme to decline the mortality rate among women sium on biomedical imaging (ISBI), hosted by the international skin
J Med Syst (2019) 43: 289 Page 19 of 19 289

imaging collaboration (ISIC). Paper presented at the Biomedical 65. Akram, T., Khan, M. A., Sharif, M., & Yasmin, M., Skin lesion
Imaging (ISBI 2018), 2018 IEEE 15th International Symposium segmentation and recognition using multichannel saliency estima-
on. tion and MSVM on selected serially fused features. Journal of
59. Navarro, F., Escudero-Vinolo, M., and Bescos, J., Accurate Ambient Intelligence and Humanized Computing, 1–-20, 2018.
segmentation and registration of skin lesion images to evaluate 66. Sultana, N. N., Mandal, B., and Puhan, N., Deep residual network
lesion change. IEEE Journal of Biomedical and Health with regularised fisher framework for detection of melanoma. IET
Informatics, 2018. Computer Vision 12(8):1096–1104, 2018.
60. Li, Y., and Shen, L., Skin lesion analysis towards melanoma detec- 67. Sarker, M., Kamal, M., Rashwan, H. A., Banu, S. F., Saleh, A.,
tion using deep learning network. Sensors 18(2):556, 2018. Singh, V. K., … Radeva, P. (2018). SLSDeep: Skin Lesion
61. Soudani, A., and Barhoumi, W., An image-based segmentation rec- Segmentation Based on Dilated Residual and Pyramid Pooling
ommender using crowdsourcing and transfer learning for skin le- Networks. arXiv preprint arXiv:1805.10241.
sion extraction. Expert Systems with Applications 118:400–410, 68. Bi, L., Kim, J., Ahn, E., & Feng, D. (2017b). Automatic skin lesion
2019. analysis using large-scale dermoscopy images and deep residual
62. Gutiérrez-Arriola, J. M., Gómez-Álvarez, M., Osma-Ruiz, V., networks. arXiv preprint arXiv:1703.04197.
Sáenz-Lechón, N., & Fraile, R. (2017). Skin lesion segmentation 69. Afza, F., Khan, M. A., Sharif, M., and Rehman, A., Microscopic
based on preprocessing, thresholding and neural networks. arXiv skin laceration segmentation and classification: A framework of
preprint arXiv:1703.04845. statistical normal distribution and optimal feature selection.
63. Bi, L., Kim, J., Ahn, E., Kumar, A., Fulham, M., and Feng, D., Microscopy Research and Technique., 2019. https://doi.org/10.
Dermoscopic image segmentation via multi-stage fully 1002/jemt.23301.
convolutional networks. IEEE Trans. Biomed. Eng 64(9):2065–
2074, 2017a.
64. Waheed, Z., Waheed, A., Zafar, M., & Riaz, F. (2017). An efficient
machine learning approach for the detection of melanoma using Publisher’s Note Springer Nature remains neutral with regard to jurisdic-
dermoscopic images. Paper presented at the Communication, tional claims in published maps and institutional affiliations.
Computing and Digital Systems (C-CODE), International
Conference on.