Journal of King Saud University – Computer and Information Sciences 34 (2022) 520–525

Contents lists available at ScienceDirect

Journal of King Saud University –

Computer and Information Sciences
journal homepage: www.sciencedirect.com

An improved bag of dense features for skin lesion recognition

Pawan Kumar Upadhyay ⇑, Satish Chandra
Department of Computer Science Engineering & Information Technology, Jaypee Institute of Information Technology, Noida, India

a r t i c l e i n f o a b s t r a c t

Article history: Skin is the largest and fastest growing organ in the human body. There are various types of skin lesions in
Received 30 September 2018 which malignancy are non-invasively detected and recognized based on their local and global attributes
Revised 4 February 2019 of the image using an image-guided system. In this work, Gradient Location and Orientation Histogram
Accepted 17 February 2019
and color features are fused together to construct the Inherently Hybrid Image Descriptor for skin lesion
Available online 25 February 2019
classification. The features obtained from these descriptor are combined to form a bag of visual words.
The improved bag is used to categorize the skin lesion classes as malignant or benign using Support
Keywords:
Vector Machine. The performance of the proposed method has been found considerably better than
Gradient Location and Orientation
Histogram (GLOH)
the current state-of-art. It also simplifies the process of diagnosis for undeclared abnormalities in the skin
Hybrid Image Descriptor (IHID) region.
Bag of visual word (BoVW) Ó 2019 The Author(s). Published by Elsevier B.V. on behalf of King Saud University. This is an open access
Scale Invariant Feature Transform (SIFT) article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
Support Vector Machine (SVM)

1. Introduction eye during examination by the medical expert (Wettschereck

et al., 1997; Armengol, 2011). In addition to this, an improvement
Skin lesion screening is a non-invasive imaging method which in each phase of Computer-Aided Diagnostic System (CAD), helps
helps in the diagnosis of skin lesion without any surgical interven- to avoid a histological analysis of skin lesions thereby reducing
tions. The skin lesions are primarily categorized into two major the rate of biopsies (Cliff, 2014; Rosendahl et al., 2012;
classes as malignant and benign. The detection and accurate clas- Mikolajczyk et al., 2010).
sification of a lesion is the primary concern for skin lesion recogni- In this paper, we give an overview of the recent development of
tion system. The severity of lesion class depends on the perceptual feature-based skin lesion recognition system. The contribution of
features of the skin, commonly available in stratum corneum (on this work is to devise a set of dermoscopic key features for recog-
the surface layer) (Iyatomi et al., 2008; Celebi et al., 2007; nition of skin lesion samples into malignant or benign.
Ballerini et al., 2013). The most severe class of pigmented lesion This paper is organized as follows. Section 2 presents an over-
is melanoma and the maximum number of cases related to mela- view of existing methods used for different classes of skin lesion
nocytic lesion occurs in cold countries. In addition to this, non- detection and classification. Some improved strategies, as well as
melanoma class of cancer is also increasing in around the world extensions as the feature vector, are also discussed in this section.
(Iyatomi et al., 2008; Ballerini et al., 2013). The various shades of Section 3 investigates common dermoscopic structures in each
color and repetitive elements of dermoscopic structure help to group of lesion. It helps to validate the reason of merging of skin
visualize the in-depth structures of a lesion in high-resolution lesion classes into two different categories as malignant or benign.
images, which are quite difficult to perceive through the naked Section 4 describes proposed method of feature bag fusion which is
used further for skin lesion classification. Finally, Section 5
describes the experimental results and its analysis which help to
⇑ Corresponding author. justify the proposed method.
E-mail addresses: [email protected] (P.K. Upadhyay), satish.chandra@
jiit.ac.in (S. Chandra).
2. Related works
Peer review under responsibility of King Saud University.

A medical image usually shows complex feature space distribu-

tions because of the two precise reasons as intra-class variability
and inter-class ambiguity. The diagnosis of pigmented skin lesions
Production and hosting by Elsevier
using dermoscopy follows certain standard rules. The fundamental

https://doi.org/10.1016/j.jksuci.2019.02.007
1319-1578/Ó 2019 The Author(s). Published by Elsevier B.V. on behalf of King Saud University.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
 P.K. Upadhyay, S. Chandra / Journal of King Saud University – Computer and Information Sciences 34 (2022) 520–525 521

rule of melanoma detection is ABCD (Asymmetry, Border, Color, vectors using BoVW techniques. The dense feature bag consists
and Diameter) which helps to develop the structural differences of two of feature descriptor having maximum level of correlation.
between the classes of skin melanoma with certain color variega- This can be achieved by local descriptor which extraction the pixel
tion (Wadhawan et al., 2011; Mikos et al., 2012). Other than that, level information from an image as two of its constituents as multi-
medical experts and researchers develop some other standards level pigmentation and structural features from the key points
methods (e.g. CASH (Bay et al., 2008) which employs the 4- (edges, corner andregion) and generates dermoscopic patterns
points and 7-points checklist, is used to classify a skin melanoma (Moreno et al., 2009; Moreels and Perona, 2007).
as malignant or benign. Although, there are standard methods for
melanoma cancer but other skin cancers recognition are still an 3. Structured classes of skin lesion
unreached area for CAD system. Based on our knowledge, there
are only two proposals in which melanocytic and non- The standard dataset of high resolution color images are cap-
melanocytic skin lesion classification performed. These can be han- tured from high definition camera. To best of our knowledge, they
dled nicely by feature descriptor based methods for skin lesion are gold standard image library (Ballerini et al., 2013) of dermofit.
classification (Iyatomi et al., 2008; Celebi et al., 2007; Ballerini There are various classes of skin lesion which are group into benign
et al., 2013; Barata et al., 2013). or malignant are based on lesion attributes as illustrated in the
These aspects can be handled nicely either by providing the Table 1, with the certain attributes of lesion which follows the laws
complete (per pixel) information of image. The spatial information of nature and benign seems to be symmetrical but malignancy sup-
of a pixel in image space is described by the descriptors. There are ports asymmetrical in terms of pattern and color.
variety of descriptors which represents global and local features of The related images of class sample which are grouped into
image space. The global features vector correspond to complete malignant or benign are shown below:
image and local features are from sub region of image space. The Group structure of skin lesion: The standard image library sam-
descriptors which are used to describe the melanoma elements ples are gold standard i.e. each sample of it, is approved by der-
(dermoscopic structures) and it pigmentation at each point of mopathologist as well as dermatologist. There are no certain
image space are formed by the combination of color descriptor clues for premalignant or nevus classes which can verify them to
and SIFT descriptor and form a Color-SIFT fused descriptor be completely malignant or benign. In this paper, Out of the ten
(Moreno et al., 2009; Salahat and Qasaimeh, 2017; Ahnlide et al., classes of standard library, we have selected only six classes of skin
2016). In this descriptor, SIFT used to mark the number of edge lesions which are kept into any one of the group as shown in Fig. 1.
of lesion and color used to mark the pigmented substances of mel- This is because the remaining classes (Actinic keratosis (AK), Seb-
anoma (Moreels and Perona, 2007; Salahat and Qasaimeh, 2017). If orrheic keratosis (SK) and Intraepithelial Carcinoma (IEC)) requires
the numbers of elements (structure or color) are more than one degree of malignancy which is not the exact scale of recognition as
then, it is considered to be chaos of structures or colors. It can be malignant or benign. In addition to this, The class ME (nevus) can
verified by related clues for each of the chaos (Epstein, 1985; never be malignant and considered to be benign when the sample
Mikolajczyk et al., 2005). is suspicious and other classes such as AK, SK, IEC are all premalig-
Before feature extraction, various steps related to dermoscopic nant always. The group selection of samples are based on certain
image processing as image enhancement and segmentation have attributes which are described below in Fig. 1.
been accounted by several authors. Enhancing the image includes
color calibration with certain normalization (Iyatomi et al., 2011; 4. Detection of dermoscopic features
Schaefer et al., 2011; Rigel et al., 2010). In addition to this, segmen-
tation techniques comprises of manual, semi-automated and auto- In this section, the authors describe the procedure of extraction
mated methods used for the lesion class and its associated for low level key descriptors which are used to identify the clues
patterns. Once, the segmentation is performed on image, its related for chaotic dermoscopic patterns of skin lesion with various shades
key features are extracted from it which represents the color and of color (Epstein, 1985; Mikolajczyk et al., 2005). The low level fea-
structure of dermoscopic images. Each Image descriptor helps to tures which are obtained from GLOH and HSV color descriptor are
transform a single value feature to feature vector in given image used to recognize malignant or benign behavior of skin lesion.
space and helps to improve its classification accuracy.
There are a large variety of textural descriptors which includes: 4.1. Dermoscopic structural features using GLOH descriptor
Gabor filter, HAAR wavelet, Gray Level Co-Occurrence Matrix
(GLCM) and shape descriptors are used in active shape model The computations for the GLOH log polar histogram are done by
(ASM) for any active instance (Amira et al., 2014; Salahat and Eqs. (1) and (2). It computes two of the component at point (x, z) of
Qasaimeh, 2017). The most frequently used color of descriptors an image, as their magnitude is the hypotenuse in Eq. (1) and the
are color moments, color histograms (Barata et al., 2014; Situ angle is the arctangent in Eq. (2) from the Cartesian coordinate of
et al., 2008; Sivic and Zisserman, 2003). Along with color, textural standard HOG are given below:
descriptors is also important as speedup robust descriptor (SURF), qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
is used to discriminate different type of melanoma patterns pðx; yÞ ¼ ðIðx þ 1; zÞ  Iðx  1; zÞÞ2 þ ðIðx; z þ 1Þ  Iðx; z  1ÞÞ2
(Catarina et al., 2013; Amira et al., 2014; Ramiro and Bykbaev,
2012). The present findings seem to be consistent with other ð1Þ
research which follows the same principal of image analysis are
performed in three major steps: (i) the borders identification and
Table 1
detection (Fix and Hodges, 1989) (ii) image features are extracted Attributes of skin lesion as benign or malignant.
from region of interest (Bay et al., 2008; Celebi, 2009; Barata et al.,
2015) (iii) and, evaluate these features with pre-calculated features Lesion Attributes Benign Malignant

of each class of skin lesion and perform skin lesion classification Dermoscopic Symmetrical in Asymmetrical in
using classification method. In order to obtained better classifica- Structures patterns patterns
Dermoscopic colors Symmetrical in color Asymmetrical in color
tion accuracy, we are targeting for more dense bag of feature
522 P.K. Upadhyay, S. Chandra / Journal of King Saud University – Computer and Information Sciences 34 (2022) 520–525

Group/Skin A:Sample Class1 B: Sample Class2 C:Sample Class 3

lesion Class

Benign

Malignant

Fig. 1. Group samples of skin lesions as Benign: (A) Pyogenic Granuloma (PYO), (B) Haemangioma (HAEM), (C) Dermatofibroma (DF) Malignant: (A) Basal Cell Carcinoma
(BCC), (B) Squamous Cell Carcinoma (SCC), (C) Malignant Melanoma (ML).

hðx; yÞ ¼ tan1 fðIðx; z þ 1Þ  Iðx; z  1ÞÞ=ðIðx þ 1; zÞ  Iðx  1; zÞÞg 5. Proposed method

ð2Þ
The proposed method consists of three phases as describe above
Log polar format is added to the SIFT descriptor which helps to in Fig. 2: (1) formation of IHID, used to describe the dermoscopic
enhance the accuracy of descriptor and amplify the rotational attributes of lesions (2) formation of BoVW (3) SVM classifier used
invariance (Ramiro and Bykbaev, 2012; Zhou et al., 2009). GLOH to classify the word index of lesion into two categories.
behave similar to the SIFT-HOG having histogram values of gradi-
ent magnitude and direction. GLOH are essentially represented as (1) IHID formation:
3D histogram. It is represented as 128D (8  4  4) in Cartesian
co-ordinates and converted to polar coordinate system by binning GLOH features are extracted from the patches of gray scale
the components as a gradient magnitude and its direction. It helps version of color image and unable to preserve the chromatic sal-
to space the skin lesion region into 17 bins with 8 way direction of iency (distinctiveness in chromatic regions) of skin lesion. For
gradient. The radically distributed bins are able to sum the gradi- converging the chrominance attribute of image pixels in isolu-
ent information in polar coordinates and capitulates 272 bin his- minance space, it is required to map the RGB color space model
togram. GLOH descriptor generate considerably important to HSV color space. The obtained color features explicitly repre-
features for gradient information which is ignored in SIFT descrip- sent the color saliency along with the structural feature descrip-
tor. Another key feature for skin lesion recognition is color and it is tor. HSV color component are computed at the same points
computed with the help of HSV color descriptor are discuss in next where GLOH features has been detected. The new image
subsection. descriptor is formed by the fusion of two features vector
extracted from the each pixel of the image patch by following
4.2. Pigmented skin lesion extraction using color features the algorithm 1.

Color descriptor extract set of color features which describe the

color properties inside the lesion. A general approach is to selected
color space that have a relation with the human perception of
vision or that are biologically inspired. The selected space charac-
terize the colors model in a way similar to that of the human mind,
namely each color is characterized by a Hue, Saturation, and Value.
The HSV color model are perceptually uniform and device depen-
dent and help to process the dermoscopy images more keenly.
Most popular method for skin lesion detection and recognition is
one dimension color descriptor which give 1D histogram. In HSV
color descriptor, the complete color space is uniformly quantized
into HSV color space with the dimension of 192, and quantization
represents each channel as 12 hues (H), 4 saturations (S) and 4
intensities (V) respectively. The quantized space encapsulate as
histogram bins having 192 dimension vector represent color fea-
tures from the image grid patch of size (4  4). It depicts that
Hue (change in color) parameter is varying with orientation and
other two parameters as Saturation (color depth) and Value (cen-
tral axis) are moving in the similar direction. Fig. 2. Proposed Method.
 P.K. Upadhyay, S. Chandra / Journal of King Saud University – Computer and Information Sciences 34 (2022) 520–525 523

Algorithm 1: Proposed descriptor where k represents to number of clusters or visual words.

3. Finally, the mapping of every descriptor to the nearest visual

1. Input: Image Im
word is based on the following equation given below
2. VDHSV(Im) <- u
3. RGLOH <- Detected GLOH Features from Key-points(Im)
VWðVDIHIDn Þ ¼ argminDistVW2VV ðVW; VDIHID n Þ ð5Þ
4. for single feature point r in RGLOH do
5. VDHSV(r) <- u
6. VD(r) <- extract image patch(r) Here, VW(VD IHID n) represents the visual word assigned to nth
7. for each s in VD(r) do descriptor and the distance argminDist (VW, VD IHID n), signifies
8. VDGLOH(s) <- extract GLOH features (s) the distance between the descriptor
9. VDHSV(s) <- extract HSV color feature(s) VDIHID n and visual word VW.
10. VDIHID(s) <- (VDGLOH(s) uVDHSV(s)) The clustering is required to reduce the dimensionality of
11. end for feature vector and represent the feature space in compact form.
12. VDIHID = VDIHID(s) The final representation of BoVW model is in a form of his-
13. output: VDIHID togram which gives the distribution of visual words. The count
of bins in histogram is equal to number of visual words in a
dictionary (i.e. k). The size of dictionary for individual or fused
The proposed descriptor (VD IHID) extracts a set of fused features bag of features in the current system is considered to be 192
locally from image patches and these are considered to be key fea- (12  4  4), which is based on the length of HSV color
ture vectors for skin lesions, as shown in the Fig. 3. These features descriptor.
are obtained from dissimilar patches of skin lesion images and it
signifies that the region of interest as lesions possess different (3) Skin lesion classifier:
shades of color and dermoscopic structures or not. This kind of
information would be watered down or even missed, if a global The number of rows represents the total number of image
representation of features was considered. samples. The selection of image samples is based on the class
having minimum number of sample i.e. 24 and total number
(2) Representation of IHID as BoVW of sample are 144 samples out-of 826 from six different classes.
There are ten different classes in standard image library out-of
Fused descriptor for skin lesion is modeled into BoVW and it them only six classes of lesion are considered for input because
describes the depicted contents of skin lesion of six different remaining classes belongs to pre-malignant. These pre-
classes. The steps for BoVW are as follows: malignant classes requires degree of malignancy requires more
clinical measures to computer, so we remove from the input
1. In BoVW technique, an input image (I) is represented as a set of set of data samples. Then the considered samples (144) are par-
image descriptors as describe in Equation (3) titioned into training (103) and testing (43) randomly. The
obtained bag of features from six different class samples of skin
I ¼fVD IHID1; VD IHID2; VD IHID3;    VD IHID i g ð3Þ lesion is classified into malignant or benign using SVM. SVM is a
discriminative classifier that learns a decision boundary that
where i denote the total number of image descriptors maximizes the margin between the classes. It is able to manage
the visual polysemy of key features for skin lesion recognition.
2. To reduce the dimensions of IHID vectors, an unsupervised clus- RBF kernel with chi -squared distance between the histograms
tering technique known as k-means is imply on the extracted is able to discriminate the skin lesion classes as malignant or
set of features to locate the cluster centers that form a visual benign more accurately. SVM is a linear classifier, for non-
vocabulary (VV) are describe as follows: linearity, it is to be transformed to other space using nonlinear
  kernel. v2 RBF kernel of SVM is computationally fast and accu-
VV ¼ VW1; VW2; VW3 ; VW4; VW5 ; VW6;    ::VWk ð4Þ rate for BoVW method.

Fig. 3. Skin lesion diagnosis using proposed method.

524 P.K. Upadhyay, S. Chandra / Journal of King Saud University – Computer and Information Sciences 34 (2022) 520–525

Recognition of lesions by SVM were done in following steps: Table 3

Confusion matrix: GLOH descriptor.

a) The size of the training set is considered to be 103  192. It Group/skin lesion Malignant Benign
represents that each image samples has fixed size of visual Malignant 0.36 0.64
vocabulary, which is represented as histogram bins. Benign 0.33 0.67
b) The histograms bins are regularizes with L1 optimisers, in
order to account the equal set of features in fused bag and
helps to remove redundant feature space. Table 4
c) More complex kernel functions as Radial basis function Confusion matrix: HSV color descriptor.
(RBF) have been used to model non-linear decision bound- Group/skin lesion Malignant Benign
aries are as follows:
Malignant 0.70 0.30
Benign 0.17 0.83
KdðRBFÞðx;yÞ ¼ e1=cdv2ðx;yÞ ð6Þ

where dv2(x; y) can be chosen to manage the intra class variegation Table 5
in feature space of distinct samples of skin lesion. The v2 distance Confusion matrix: proposed descriptor.
consider to be better distance metrics when comparing the his- Group/skin lesion Malignant Benign
togram structures of skin lesions in BoVW method. Malignant 0.72 0.28
Benign 0.17 0.83

6. Results and discussions

Performances of fused bag of features: Table 5 shows the auto-

A generalized BoVW method is used for skin lesion diagnosis
mated recognition rate (ARR) for finest combinations among the
having fixed length of vocabulary for each image samples using
two features. Best result 78% accuracy is achieved by combining
single or fused descriptors. The experiments are performed on
GLOH and HSV Color feature. This validates the usefulness of the
the platform of MATLAB-2014b with system configuration of Dual
color feature descriptor for skin lesion recognition. Table 5 shows
Core I-5 Processor having 8 GB RAM and GPU of 920 M series with
the confusion matrix of fused features which leads to a enhance-
the size of 2 GB.
ment in skin lesion classification. The results are shown in Table 3
Parametric evaluation for dense Bag-of-features: As mentioned
for GLOH descriptor which detect malignancy with the recognition
above the dense features are able to detect the symmetry of der-
rate of (RR = 36%). Furthermore, the results are improved when the
moscopic key feature as dermoscopic structures and colors and
classification was performed on fused features (after fusion with
able to recognize the malignant or benign samples of skin lesions.
color feature) as shown in Table 5 with the recognition rate of
Following performance measures are considered for the lesion
malignant (RR = 72%) and benign (RR = 83%). The weighted values
classification: (a) No of clusters, (b) No of features (c) No of itera-
are compared in confusion matrix and it shows that the color fea-
tions or time to converge for iteration to generate efficient visual
ture vector is dominating the key feature vector obtained from the
bag of hybrid features. The basic attributes for the computation
bag of IHID descriptor.
of descriptor are (a) Strong features are selected from each of the
Benign/Malignant is considered to be a false negative rate; our
bag and keeping the size of the vocabulary is fixed i.e.192. (b)
objective is to improve the true positive rate in the confusion
The key point selection is based on constant radius for each of
matrix as describe in Table 5 of bag of proposed descriptor. The
the descriptor. (c) So, the feature vector which are encode in a pro-
obtained features in bag of HSV color descriptor and bag of pro-
posed bag of visual word is 144  192.
posed descriptor were same for fixed vocabulary (192) as shown
Results of the proposed method is describes below in the fol-
in Table 2, is the main reason of constant false negative rate.
lowing Tables 3–5 as given below:
For dermoscopic skin lesion recognition, the mean accuracy of
Performances of Individual Bag of Feature: In this sub-section,
SVM classifier for GLOH descriptor (52%),HSV Color descriptor
Tables 2–4 shows the confusion matrix of individual and fused fea-
(76%) and IHID descriptor (78%) is describe above in the Fig. 4. In
tures which are obtained by evaluating these features on SVM clas-
addition to this, they obtained sensitivity, specificity and accuracy
sifier and compute averaged run time after splitting the data in the
ratio of [T(70):T(30)].
The values along the diagonal of the matrixes represent the
recognition rate of samples corresponds to each class, and the
numbers other then diagonal show the error rate (misclassification
rate) denoted as error rate (ER = 17). We can see that color feature
perform well for classes with very distinguishable color such as
benign (RR = 83%) and malignant (RR = 70%) in Table 4. However,
the GLOH features are not able to distinguish well between benign
(RR = 67%) and malignant (36%) as shown in the Table 3.

Table 2
Comparison of descriptors with fixed size of vocabulary.

Descriptors used to No of clusters No of No of iterations/

describes skin lesions (K)/size of features time to converge
Vocabulary
GLOH 192 2022 7/0.94 s/iteration
HSV Color 192 181,656 48/0.94 s/iterations
IHID 192 181,656 51/0.85 s/iterations
Fig. 4. Comparison of proposed descriptor with conventional descriptor using SVM.
 P.K. Upadhyay, S. Chandra / Journal of King Saud University – Computer and Information Sciences 34 (2022) 520–525 525

Table 6 References
Comparison of proposed method with the current State-of -the-art.

Authors Method No of Accuracy Ahnlide, I., Bjellerup, M., Nilssom, F., Nielsen, K., 2016. Validity of ABCD rule of
dermoscopy in clinical practice. Acta Derm. Venereol. 96, 367–372.
classes (%)
Amira, B.M., Najjar, A., Zagrouba, E., 2014. Image flower recognition based on a new
Ballerini et al. Color and texture, K-NN 5 74 method for color feature extraction. In: International Conference on Computer
(2013) Vision Theory and Applications (VISAPP 2014), pp. 201–206.
Situ et al. (2008) Color histogram, Gabor filter, 2 82.2 Armengol, E., 2011. Classification of melanomas in situ using knowledge discovery
BoVW, K-NN with explained case-based reasoning. Artif. Intell. Med. 51, 93–105.
Mikos et al. (2012) GLCM, PNN Neural network 2 69.5 Ballerini, L., Fisher, R.B., Aldridge, B., Rees, J., 2013. A color and texture based
Wadhawan et al. Color histogram, Haar wavelet, 2 76.4 hierarchical K-NN approach to the classification of non-melanoma skin lesions.
Color Med. Image Anal., 63–86
(2011) SVM
Barata, C., Marques, J., Mendonca, T., 2013. Bag-of-features classification model for
Proposed method GLOH, HSVcolor, BoVW, SVM 6 78
the diagnose of melanoma in dermoscopy images using color and texture
descriptors. Int. Conf. Image Anal. Recogn., 547–555
Barata, C., Ruela, M., Francisco, M., Mendonça, T., Marques, J.S., 2014. Two systems
for the detection of melanomas in dermoscopy images using texture and color
features. IEEE Syst. J. 8 (3), 965–979.
for two class of lesion are also describes in the Fig. 4. The accurate Barata, C., Celebi, E.M., Marques, S.J., 2015. Melanoma detection algorithm based on
detection of skin lesion class is crucial for successful treatment. feature fusion. In: 37th Annual International Conference of the IEEE Engineering
in Medicine and Biology Society, pp. 2653–2656.
However, there are various methods for skin lesions recognition Bay, H., Ess, A., Tuytelaars, T., Gool, V.L., 2008. Speeded-Up Robust Features (SURF).
but the proposed method has the potential of enhancing the cur- Comput. Vis. Image Underst. 110 (3), 346–359.
rent clinical paradigm in the domain of dermatology. Catarina, B., Jorge, S.M., Rozeira, J., 2013. Evaluation of color based key points and
features for the classification of melanomas using the bag-of-features model.
Comparison with a Current-state-of-art: We compared our work Advances in Visual Computing: 9th International Symposium, Rethymnon,
with lesion recognition methods proposed by several researchers Crete, Greece, pp. 40–4926.
in the domain of skin lesion classification and recognition. A direct Celebi, M.E., 2009. Lesion border detection in dermoscopy images. Comput. Med.
Imag. Graph. 33 (2), 148–153.
comparison is not possible because of different dataset and valida-
Celebi, M., Kingravi, H., Uddin, B., Iyatomi, H., Aslandogan, Y., Stoecker, W., Moss, R.,
tion measures used in these methods. The results are shown in 2007. A methodological approach to the classification of dermoscopy images.
Table 6. Comput. Med. Imaging Graph. 31 (6), 362–373.
In the work, we identifies the simplified criteria of skin lesion Cliff, R., 2014. Dermotoscopy chaos and clues: a decision algorithm for pigmented
skin lesion. Practical Dermoscopy., 23–28
diagnostic and help to boost its recognition rate (RR) up to 78%. Epstein, E., 1985. Magnifiers in dermatology: a personal survey. J. Am. Acad.
The proposed method uses bag of visual words (BoVW) technique Dermatol. 134, 687–698.
which help to identify the clues for chaotic skin lesion using der- Fix, E., Hodges, J.L., 1989. Discriminatory analysis. Nonparametric discrimination:
consistency properties. Int. Stat. Rev. 57 (3), 238–247.
moscopic structures and colors features and discriminate them, Iyatomi, H., Oka, H., Celebi, M.E., Hashimoto, M., Hagiwara, M., Tanaka, M., 2008. An
to be malignant or benign. In the table (Situ et al., 2008), a compar- improved internet-based melanoma screening system with dermatologist-like
ison is drawn between similar technique having identical proce- tumor area extraction algorithm. Comput. Med. Imaging Graph. 32 (7), 566–
579.
dure of skin lesion recognition. These procedures were based on Iyatomi, H., Celebi, M.E., Schaefer, G., Tanaka, M., 2011. Automated color calibration
feature analysis of dermoscopic image used to describe the pat- method for dermoscopy images. Comput. Med. Imag. Graph. 35 (2), 89–98.
terns of skin lesion or its colors. In fact, the complexity of this algo- Mikolajczyk, K., Tuytelaars, T., Schmid, C., Zisserman, A., Matas, J., Schaffalitzky, F.,
Kadir, T., Van, L., 2010. Affine covariant features. http://www.robots.ox.ac.uk/
rithm is O(V*D*P) where V is the vocabulary size (number of visual ~vgg/research/affine, 57–59.
words), D is the dimension of the feature and P is the number of Mikolajczyk, K., Tuytelaars, T., Schmid, C., Zisserman, A., Matas, J., Schaffalitzky, F.,
detected points. Kadir, T., Gool, V.L., 2005. A comparison of affine region detector’s. Int. J.
Comput. Vis. 65 (2), 43–72.
Mikos, E., Sioulas, I., Sidiropoulos, K., Cavouras, D., 2012. An android-based pattern
recognition application for the characterization of epidermal melanoma. In:
7. Conclusions Workshop on Biomedical Instrumention and Related Engineering, Athens. pp.
63–72.
Moreels, P., Perona, P., 2007. Evaluation of features detectors and descriptors based
This work has shown that GLOH and color features were fused on 3d objects. Int. J. Comput. Vis. 73 (3), 263–284.
together to construct the new descriptor IHID for skin lesion recog- Moreno, P., Bernardino, A., Victor, S.J., 2009. Improving the SIFT descriptor with
nition. The features obtained from these descriptor are represented smooth derivative filters. Pattern Recogn. Lett. 30 (1), 18–26.
Ramiro, J.F., Bykbaev, V.R., 2012. Image retrieval based on the combination of RGB
in a form of bag of visual words, which was used to categorize the and HSV’s histograms and Colour Layout Descriptor. pp.1-10, DOI:10.17163/
skin lesion classes as malignant or benign using Support Vector ings.n7.2012.01
Machine. The proposed approach performed screening on six dif- Rigel, D.S., Russak, J., Friedman, R., 2010. The evolution of melanoma diagnosis: 25
years beyond the ABCDs. Cancer J. Clin. 60 (5), 301–316.
ferent class samples of skin lesion, without any incident of projec- Rosendahl, C., Cameron, A., McColl, I., Wilkinson, D., 2012. Dermatoscopy in routine
tion, scale space and any lighting condition. Furthermore, it is fast, practice ‘chaos and clues. Aust. Fam. Physician 41 (7), 482–487.
accurate and cost-effective approach which leads to a medical Salahat, E., Qasaimeh, M., 2017. Recent advances in features extraction and
description algorithms: a comprehensive survey. arXiv:1703.06376v1 [cs.CV].
description of lesion in simplest way with the average accuracy
Schaefer, G., Rajab, M.I., Celebi, M.E., Iyatomi, H., 2011. Color and contrast
of 78%. The results obtained in the proposed approach are quite enhancement for improved skin lesion segmentation. Comput. Med. Imag.
promising as compared to current state-of-art. In future, the pro- Graph. 35 (2), 99–104.
posed approach may be applied to different modalities of medical Situ, N., Yuan, X., Chen, J., Zouridakis, G., 2008. Malignant melanoma detection by
bag-of-features classification. Int. Conf. IEEE Eng. Med. Biol. Soc., 3110–3113
images related to many other vital organs. Furthermore, some Sivic, J., Zisserman, A., 2003. Video Google: a text retrieval approach to object
modifications will be required in the proposed method so that matching in videos. Int. Conf. Comput. Vis. 0-7695-1950-4/03.
degree of malignancy may be considered and it will helps to recog- Wadhawan, T., Situ, N., Rui, H., Lancaster, K., Yuan, X., Zouridakis, G., 2011.
Implementation of the 7-point checklist for melanoma detection on smart hand
nize the premalignant class samples of skin lesion. held devices. Int. Conf. IEEE Eng. Med. Biol. Soc., 3180–3183
Wettschereck, D., Aha, D.W., Mohnri, T., 1997. A review and empirical evaluation of
feature weighting methods for a class of lazy learning algorithms. Artif. Intell.
Rev. 11, 273–314.
Conflict of interest
Zhou, H., Chen, M., Rehg, J.M., 2009. Dermoscopic interest point detector and
descriptor. In: IEEE International Symposium on Biomedical Imaging. https://
The authors declare that they have no competing interests. doi.org/10.1109/ISBI.2009.5193307.