IEEE SYSTEMS, MAN AND CYBERNETICS SOCIETY SECTION

Received 5 April 2023, accepted 6 May 2023, date of publication 10 May 2023, date of current version 30 May 2023.
Digital Object Identifier 10.1109/ACCESS.2023.3274848

DeepSkin: A Deep Learning Approach for Skin

Cancer Classification
H. L. GURURAJ 1 , (Senior Member, IEEE), N. MANJU 2 , A. NAGARJUN2 ,
V. N. MANJUNATH ARADHYA3 , AND FRANCESCO FLAMMINI 4 , (Senior Member, IEEE)
1 Departmentof Information Technology, Manipal Institute of Technology Bengaluru, Manipal Academy of Higher Education, Manipal 576104, India
2 Departmentof Information Science and Engineering, JSS Science and Technology University, Mysuru 570006, India
3 Departmentof Computer Application, JSS Science and Technology University, Mysuru 570006, India
4 IDSIA USI-SUPSI, University of Applied Sciences and Arts of Southern Switzerland, 6928 Manno, Switzerland

Corresponding author: Francesco Flammini ([email protected])

ABSTRACT Skin cancer is one of the most rapidly spreading illnesses in the world and because of the limited
resources available. Early detection of skin cancer is crucial accurate diagnosis of skin cancer identification
for preventive approach in general. Detecting skin cancer at an early stage is challenging for dermatologists,
as well in recent years, both supervised and unsupervised learning tasks have made extensive use of deep
learning. One of these models, Convolutional Neural Networks (CNN), has surpassed all others in object
detection and classification tests. The dataset is screened from MNIST: HAM10000 which consists of seven
different types of skin lesions with the sample size of 10015 is used for the experimentation. The data
pre-processing techniques like sampling, dull razor and segmentation using autoencoder and decoder is
employed. Transfer learning techniques like DenseNet169 and Resnet 50 were used to train the model to
obtain the results.

INDEX TERMS Skin cancer, segmentation, deep learning, CNN, Densenet169, Resnet50.

I. INTRODUCTION technique dermoscopy uses a magnified and lighted picture

A tumor is formed when healthy cells begin to change of the affected skin area to increase visibility of the spots,
and grow out of control. Both cancerous and noncancerous therefore reducing facial reflection [4]. Skin cancer early
tumors are conceivable. Malignant tumors are those that have detection is still a prized possession. It’s difficult to tell if
the potential to grow and spread to other areas of the body [1]. a skin lesion is benign or malignant because they all seem
A benign tumor may form, but it does not usually spread. Skin similar. The sun’s harmful ultraviolet (UV) rays and the usage
cancer is the result of abnormal skin cell growth. It is the most of UV tanning beds are the two most common causes of
prevalent cancer nowadays and occurs everywhere. Every skin cancer. It is particularly difficult for dermatologists to
year, various forms of melanomas are thought to cause more distinguish between melanoma and non-melanoma lesions
than 3.5 million cases to be discovered [2], [3]. This number because of the low degree of difference between lesions and
exceeds the sum of cases of lung, bone, and colon cancers. skin [5]. The main problem of similar opinion is largely
In reality, a person with melanoma dies every 57 seconds. dependent on private judgment and is scarcely reproducible.
When cancer is detected in dermoscopy images in advance, With the help of robotization using operation of deep literacy
the survival percentage is significantly boosted. Therefore, helps the case to get the early opinion report and grounded on
accurate automatic skin excrescence discovery will undoubt- the report case can consult dermatologists for treatment [6].
edly help pathologists become more skilled and productive. An early diagnosis of skin cancer is crucial and has limited
The purpose of the dermoscopy technique is to improve each number of available treatment options. Accurate evaluation
melanoma patient’s performance. Noninvasive skin imaging and the capacity to accurately identify skin cancer are critical
components of a skin cancer prevention approach. Even in
The associate editor coordinating the review of this manuscript and literacy tasks that are unsupervised, deep literacy has been
approving it for publication was Huiyan Zhang . widely adopted [7]. Object detection and bracket tasks have
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
VOLUME 11, 2023 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ 50205
 H. L. Gururaj et al.: DeepSkin: A Deep Learning Approach for Skin Cancer Classification

been dominated by Convolutional Neural Networks (CNN). Some of the most commonly used machine learning tech-
As a result, trained end-to-end in a controlled environment, niques were employed. 328 images of benign melanoma and
CNNs eliminate the need for humans to manually create fea- 672 images of melanoma were retrieved from the ISIC collec-
ture sets. The use of Convolutional Neural Networks (CNNs) tion. With SVM classifiers, a 97.8% accuracy and 0.94 area
to categorize lesions in skin cancer in recent years has out- under the curve resulted in a classification [2], [12]. Further-
performed skilled mortal specialists. more, the KNN results showed that the Sensitivity was 86.2%
and the Specificity was 85%. The proposed strategy in [13]
II. LITERATURE REVIEW focuses on the identification and categorization of skin cancer
The methodology for classification of skin lesions with through the specialised use of machine learning methodology.
fine-tuned neural networks is proposed in [8]. For the pur- The classification rate for unsupervised learning using the
pose of balancing the dataset, the skin lesion photos are k-means algorithm is 52.63%. The k-means algorithm divides
resampled. A combination of DenseNet and U net is trained the input data into n data points and k clusters. Two clusters
for segmentation and then used to fine-tune the following are produced when melanoma skin cancer is detected; one
classifiers. The encoder element of the segmentation model’s cluster is for cancer detection and the other is for non-cancer
extracted architecture is then trained to categorize the seven detection. Back Propagation Neural Network classification
different skin disorders. The classification model’s average accuracy has been found to range from 60% to 75%, whereas
balanced accuracy was 0.836 in the test set and 0.840 in the Support Vector Machine (SVM) accuracy ranges from 80%
validation set. An innovative strategy is put forth in [9] that to 90%. As a result, Support Vector Machine performs bet-
can pre-process the image automatically before segmenting ter than K-means clustering and Back Propagation Neural
the lesion. Hair, gel, bubbles, and specular reflection are just Network classification. Unsupervised learning is used as the
a few of the undesirable artefacts that the system filters. foundation for the K-means method. The proposed approach
An innovative method for identifying and painting the hairs in [14] concentrates on classification issues. The main goal
visible in cancer images is given using the wavelet idea. of this study is to categorise skin lesions using deep learning,
Utilizing an adaptive sigmoidal function that manages the especially using the CNN method. The dataset was compiled
localised intensity distribution within the pictures of a spe- using ISIC. The techniques include the employment of trans-
cific lesion, the contrast between the lesion and the skin is fer learning algorithms like Inception V3, Resnet and VGG-
improved. We then provide a segmentation method to pre- 16 and Mobilenet, as well as data augmentation, image nor-
cisely separate the lesion from the surrounding tissue. On the malisation, and image normalizing. The approaches utilised
European dermoscopic image database, the proposed method in [15] are focused on employing the supervised learning
is tested. The proposed system is focusing on classifying method to categorise skin lesions. With the specialised imple-
skin lesions in deep learning with specific implementation mentation of computer aided diagnosis, MAP estimate can
of CNN approach [10]. The methods in this paper includes carry out numerous routine activities in automated skin lesion
the screening of a dataset from MNIST: HAM10000 which diagnosis (CAD). Lesion segmentation, hair detection, and
consists of seven different types of skin lesions with the pigment network detection are some of the techniques. The
sample size of 10015. The methods in this paper included developed model has an accuracy rate of 86%. The techniques
training the model with the help of CNN and obtained an include computer-aided diagnostics and MAP estimate for
accuracy of 78%. In the study, a deep fully convolutional categorising skin lesions. The methods proposed in [16]
neural network is presented for semantic pixel-by-pixel seg- is focusing on detection and classification of skin cancer
mentation. A pixel- wise classification layer, a corresponding with specific implementation of deep learning approach. The
decoder network, and an encoder network make up the train- methods in this paper includes the screening of a dataset from
able network [1]. Convolution layers are identical to those MNIST: HAM10000 which consists of seven different types
in VGG-16’s 13 convolution layers. As its name implies, the of skin lesions with the sample size of 10015 and PH2 dataset
decoder network’s primary goal is to convert encoder feature which contains 200 images of skin lesions. The methods
maps into full input resolution feature maps. The method is include data augmentation and the model is trained using
focusing on classifying skin lesions in deep learning with deep learning architectures like mobilenet, and VGG-16 [17].
specific implementation of CNN approach [2]. The methods The accuracy obtained is 81.52% with mobilenet and 80.07%
in this paper includes the screening of a dataset from MNIST: with VGG-16.
HAM10000 which consists of seven different types of skin
lesions with the sample size of 10015. The methods in this
paper included training the model with the help of CNN III. METHODOLOGY
and obtained an accuracy of 88% and used transfer learning Detecting skin cancer at an early stage is challenging for
methods like the Resnet model. An application of machine dermatologists. With the extensive use of deep learning pro-
learning with a focus on skin cancer categorization has been cedures as shown in Figure1 helps to classify the seven types
developed [11]. Pre-processing, segmentation, feature extrac- of skin cancer images. Deep learning methods like Convo-
tion, and classification are all included in the research. The lutional Neural Networks (CNN), has surpassed all others in
ABCD rule, GLCM, and HOG were used to extract features. object detection and classification tests.

50206 VOLUME 11, 2023

H. L. Gururaj et al.: DeepSkin: A Deep Learning Approach for Skin Cancer Classification

FIGURE 2. Data visualization.

Oversampling (up sampling): Balanced datasets can

FIGURE 1. Flow of methodology.
be created by altering data classes that are not equal in
value. To compensate when there’s not enough data, the
oversampling approach increases uncommon samples’ sizes.
Minorities are more prevalent in the training set when over-
A. DATASET DESCRIPTION sampling is used. Oversampling has the benefit of retaining
The dataset contains images of skin lesions from seven every observation from the original training set for both the
different classes. HAM10000 and ISIC from the National minority and majority classes. Oversampling works very well
Institutes of Standards and Technology (NIST). It includes for small datasets. Undersampling (down sampling): This
10015 dermatoscopic images of all significant diagnostic approach reduces the size of the overabundant class in order
categories of pigmented skin lesions. Which includes actinic to balance the dataset. It aims to reduce the number of major-
keratoses and intraepithelial carcinoma/disease Bowen’s ity samples in order to get a more even distribution of classes.
(akiec), basal cell carcinoma (bcc), benign keratosis-like As a result of eliminating data from the original, it may lose
lesions (solar lentigines/seborrheic keratoses and lichen- out on important information. The bulk of the population
plan (angiomas, angiokeratomas, pyogenic granulomas and can benefit from this strategy. It decreases the number of
hemorrhage, vasc). observations from the dominating class in order to make the
data set more equal. We should use this method only when
B. DATA VISUALIZATION the dataset is big and the amount of training samples helps
The changes we make to our data before submitting it to the decrease run time and storage difficulties. Figure 3. shows the
algorithm are referred to as pre-processing. Raw data may be balanced dataset of the skin lesions to the respective classes
transformed into a clean data collection via pre-processing. they belong to.
In other words, data obtained from various sources is often
obtained in a raw format that is inappropriate for analysis. The C. REMOVING THE NOISE FROM THE SKIN IMAGE
visual presentation of information and data is known as data The Dull Razor method can be used to remove hair from skin
visualization. Figure 2. shows the graphical representation of images. Dull Razor takes the following actions:
data visualization of the skin lesions.
1. The dark hair locations are identified by the use of
The changes we make to our data before submitting it to the
a generalized grayscale morphological closing tech-
algorithm are referred to as pre-processing. Raw data may be
nique.
transformed into a clean data collection via pre-processing.
2. A bilinear interpolation is used to replace the hair pix-
In other words, data obtained from various sources is often
els that have been verified as having a thin and long
obtained in a raw format that is inappropriate for analysis. The
structure.
visual presentation of information and data is known as data
3. To smooth out the hair pixels that have been replaced,
visualization. Figure 2. shows the graphical representation of
it uses an adaptive median filter.
data visualization of the skin lesions.
The dataset is unbalanced in this case. Hence, we can use Let Gt represent the original band’s generalized grayscale
a technique called sampling to balance the dataset. Data sam- closure pictures, and Ot , S1 , S2 , and S3 represent the hori-
pling is a set of techniques for reshaping a training dataset so zontal, vertical, and diagonal structure elements. Gt can be
that classes are more evenly distributed. Once the dataset has written as given in equation (1)
been balanced, standard machine learning algorithms may be
used to train on the adjusted dataset without any adjustments. Gt = |Ot − max{Ot · S1 Ot · S2 Ot · S1 }| (1)

VOLUME 11, 2023 50207

 H. L. Gururaj et al.: DeepSkin: A Deep Learning Approach for Skin Cancer Classification

FIGURE 4. Removal of hair from the skin lesions.

D. SEGMENTATION OF IMAGE USING ENCODER AND

DECODER
FIGURE 3. Data visualization.
The technique of segmenting an image involves dividing the
foreground from the background or grouping pixels accord-
ing to how similar they are in terms of color or shape. The
where · indicates that the operation is in grayscale. Further- encoder and decoder technology are used to segment images
more, for the coordinates (x, y), Mr (x, y), the binary hair for biomedical image segmentation.
mask pixel is computed as shown in equation (2) Encoder: Convolutions and max pooling are done at the
encoder. From the original fully connected layers, 13 convo-
Mr (x, y) = 1, if Gr (x, y) > T
lutional layers were eliminated. The associated max pooling
= 0, otherwise (2) indices (locations) are stored while executing 2 × 2 max
pooling.
where T is a predefined threshold value.
For green and blue channels, we may use a phrase like that. φ:X →F
It’s the aggregate of the three-color channels’ hair masks that ψ :F →X
makes up the final mask for the original image, M as given in φ, ψ = argmin∥X − (ψ ◦ φ)X ∥2 (4)
equation (3).
In the encoder function, indicated by φ, a latent space F is
M = Mr ∪ Ma ∪ Mb (3) used to map the original data X. To get the data out of the
bottleneck, we use the decoder function, which we indicate
The hair masks for the appropriate color channels are Mr , by ψ. Here, the result is the same as what you got when you
Ma , and Mb . Due to the thresholding procedure and image put it in. After some generic nonlinear compression, we are
noise, it is difficult to determine exactly where the borders are essentially attempting to restore the original image which
in the image. It is common to produce tiny lines around the is given in equation (4). Standard neural network function
hair region’s boundary. The thin lines are smoothed out is transferred via an activation function, which represents
using an adaptive median filter in this step. The reduction the latent dimension z in the encoding network is given in
of intensity shifts from one pixel to the next is its main equation (5).
goal. In this filter, the pixel value is switched out for the
Z = σ (Wx + b) (5)
median value. By arranging all the pixel values in ascending
order, the median is calculated by substituting the calculated It is also possible to describe the decoding network this
pixel with the middle pixel value. Image smoothing with way, but with a wide range of weights, biased and activation
mean filtering minimizes the intensity variation across pixels functions. The autoencoder’s purpose is to pick our encoder
in a plain, intuitive, and straightforward manner. It is often and decoder algorithms such that we only use the lowest
used to lessen the amount of noise in images. The median quantity of data to encode the image for it to be regenerated
filter compares each pixel in the picture to its immediate on the other side.
surroundings to see if it is representative of those in the near Decoder: Upsampling and convolutions are done at the
vicinity. Instead of just replacing the pixel value with the decoder. Each pixel has a softmax classifier at the end. Prior
average of surrounding pixel values, the median of those to sampling, encoder layers remember their highest possible
values is utilized. The outcome of subtracting the closed maximum pooling indices. In the end, a K-class softmax
picture and the original image is black hat filtering. This classifier is used to predict each pixel’s class. Image thresh-
is used to extract minute details from the structures in the olding:Images are resized when they are scaled down. A wide
original image, highlighting brilliant items in dark back- range of image processing tasks benefit from scaling. As a
grounds. Figure 4. shows the removal of hair from the skin result, the number of pixels in an image may be reduced,
lesion. which has several benefits. Since the more pixels in an image

50208 VOLUME 11, 2023

H. L. Gururaj et al.: DeepSkin: A Deep Learning Approach for Skin Cancer Classification

TABLE 1. Training the model for 30 epochs.

FIGURE 5. Output of skin lesion thresholding.

there are, the more input nodes there are, which increases the
model’s complexity, training a neural network may be short-
ened. It also helps with zooming in on images. We frequently
need to resize an image, either to reduce it or to scale it up
to suit the size requirements. Figure5 shows the output for
image thresholding for skin lesion image. a neural network
may be shortened. It also helps with zooming in on images.
We frequently need to resize an image, either to reduce it or
to scale it up to suit the size requirements. Figure 5. shows
the output for image thresholding for skin lesion image.

IV. SYSTEM IMPLEMENTATION

Convolutional Neural Networks (CNNs) are built in such a
way that they can account for the input’s spatial structure.
They were originally developed to work with images and
were inspired by the visual system of the mouse. CNNs fea-
ture fewer parameters than ordinary neural networks, allow-
ing for the efficient training of very deep designs (usually
more than 5 layers which is almost impossible for fully-
connected network. An input layer, an output layer, and
several hidden layers make up a convolutional neural net-
work. Convolutional, RELU (activation function), pooling,
fully connected, and normalization layers are often seen in
a CNN’s hidden layers. Pre-trained models like Densenet169
and Resnet50 were used to make a comparison study between
oversampling and undersampling technique. Densenet169
architecture didn’t supported for Oversampling technique.
Table 1 shows training the model for 30 epochs.
Learning graph: At each iteration, the training accuracy While the input is below zero, the output is zero, but when
improves but the validation loss is detected more often, result- the input rises above a certain threshold, it has a linear rela-
ing in a reduction in the training accuracy at each iteration tionship dependent variable f(x) = max (0, x) as shown in
of the plot and Figure 8 shows the learning graph of training Figure 8.
and validation accuracy plotted for 30 epochs. Figure 6 shows Softmax is a function that takes the output values from the
the learning graph of training and validation accuracy for model and turns it into a probability function. The softmax
30 epochs with the batch size of 32 and dropout layer of 0.5 to function is given by the formula in equation (7).
avoid overfitting of the model. Figure 8 shows the learning e zi
graph of training and validation loss for 30 epochs. (7)
K
ezj
P
The accuracy of the model obtained is 91.2%. The RELU
activation functions is used for training Relu (Rectified linear j=1

Unit) shown in equation (6) Performance on issues involving sparse gradients is enhanced
It is defined as by the use of the adaptive gradient algorithm (AdaGrad) (e.g.,
natural language and computer vision problems). Addition-
y = max(0, x) (6)
ally, using Root Mean Square Propagation (RMSProp), per-
Rectified linear is a more is more interesting transform that parameter learning rates are adjusted to correspond to the
activates a node only if the input is above a certain quantity. mean of recent weight gradient magnitudes (e.g., how quickly

VOLUME 11, 2023 50209

 H. L. Gururaj et al.: DeepSkin: A Deep Learning Approach for Skin Cancer Classification

FIGURE 7. Training and validation a graph for 30 epochs.

FIGURE 6. Training and validation accuracy graph for 30 epochs.

TABLE 2. Hyperparameters used for denset 169.

it is changing). This implies that the method performs well

for online and non-stationary situations (e.g., noisy). Adam is
aware of the advantages of RMSProp and AdaGrad. Binary
cross entropy is the chosen loss function. The definition of
the loss function is given in equation (8).
N
1X
Hp (q) = − yi × log(p(yi ))+(1-yi ) × log(1 − p(yi ))
N
i=1
(8) FIGURE 8. Relu Activation function graph.
The accuracy of the model obtained is 91.2%. The training
and testing ratio were set to different ranges like 80:20, 70:30
and 40:60. For 70:30 ratios, the model is trained for 20 epochs
with a learning rate of 0.000001. The accuracy of the model
obtained is 87.7%. Figure 9 shows the training and validation
accuracy graph plot and Figure 10 shows the training and the
validation loss graph plot.
It takes the model 30 epochs to train at 0. 0001 learning
rate for 40:60 ratios. The model’s accuracy was determined
to be 82.8%. Figure 11 shows the training and the validation
accuracy and Figure 12 shows the loss graph plot. Table 1
shows the hyperparameters used for densenet 169 model.
Resnet50 architecture did not supported for undersampling
technique. The model is trained for 10 epochs with learning
rate of 0.01 as given in Table 2.
At each iteration, the training accuracy improves but the
validation loss is detected more often, resulting in a reduction FIGURE 9. Training and validation accuracy graph for 20 epochs.
in the training accuracy at each iteration of the plot as shown
in Figure 13.
The accuracy of the model obtained is 81%. To increase with a learning rate of 0.0001. Table 3 shows training the
training accuracy, the model is trained across three epochs model for 3 epochs.

50210 VOLUME 11, 2023

H. L. Gururaj et al.: DeepSkin: A Deep Learning Approach for Skin Cancer Classification

TABLE 3. Training the model for 10 epochs.

FIGURE 10. Training and validation loss graph for 20 epochs.

FIGURE 13. Learning graph of training for 10 epochs.

FIGURE 11. Training and validation accuracy graph for 20 epochs.

FIGURE 12. Training and validation loss graph for 20 epochs. FIGURE 14. Learning graph of training for 3 epochs.

The accuracy of the model obtained is 83%. Training loss Figure 15 shows the model accuracy graph plot and Figure 16
versus validation loss and training accuracy vs validation shows the model loss learning graph plot.
accuracy are shown in the graph given below as shown in For 40:60 the model is trained for 30 epochs with a learning
Figure 14. rate of 0.0001. The accuracy of the model obtained is 81.6%.
The training and testing ratio were conducted on different At each iteration, the training accuracy is shown below the
ranges like 80:20, 70:30 and 40:60. Using a learning rate of validation accuracy displayed in the plot of training loss
0.00001 and 30 training epochs, the model is trained to learn versus validation loss. Figure 17 shows the model accuracy
70:30 with the batch size of 16 dropout layer of 0.5 to avoid graph plot and Figure 18 shows the model loss learning graph
overfitting of the model. The model’s accuracy is 80.9%. plot.

VOLUME 11, 2023 50211

 H. L. Gururaj et al.: DeepSkin: A Deep Learning Approach for Skin Cancer Classification

FIGURE 15. Training and validation accuracy graph for 30 epochs.

FIGURE 18. Training and validation loss graph for 30 epochs.

FIGURE 16. Training and validation loss graph for 30 epochs.

FIGURE 19. Confusion matrix for undersampling.

FIGURE 17. Training and validation accuracy graph for 30 epochs.

V. RESULT AND DISCUSSION FIGURE 20. Confusion matrix for Oversampling.

Confusion matrix can be used to produce the F1 score as

an evaluation metric for the model. The performance of a both undersampling and oversampling techniques and obser-
classification model (or ‘‘classifier’’) on a set of test data for vations are made by using confusion matrix. The accuracy
which the real values are known can be represented using a obtained by using undersampling technique is 91.2% and f1
table called a confusion matrix. The model is trained with measure of 91.7%.

50212 VOLUME 11, 2023

H. L. Gururaj et al.: DeepSkin: A Deep Learning Approach for Skin Cancer Classification

TABLE 4. Training the model for 3 epochs. a person’s vulnerability to the sun’s UV radiation. Given
the limited resources available, early identification of skin
cancer is essential. Accurate diagnosis and identification
viability are generally essential for skin cancer prevention
strategies. Additionally, dermatologists have trouble seeing
skin cancer in its early stages. The use of deep learning for
both supervised and unsupervised applications has increased
TABLE 5. Classification report. significantly in recent years. Convolutional Neural Networks
(CNNs) are one of these models that have excelled in object
identification and classification tasks (CNN). The dataset is
filtered from MNIST: HAM10000, which has a sample size of
10015 and includes seven different types of skin lesions. Data
preprocessing methods include sampling, segmentation using
an autoencoder and decoder, and dull razor. The model was
trained using transfer learning methods like DenseNet169 and
Resnet 50. Different ratios were used for the training and
assessment, including 80:20, 70:30, and 40:60. When under-
sampling and oversampling were compared, DenseNet169’s
undersampling technique produced accuracy of 91.2% with a
f1-measure of 91.7% and Resnet50’s oversampling technique
TABLE 6. Comparison between undersampling and oversampling with
different split. produced accuracy of 83% with a f1-measure of 84%. The
future extension of this study includes increasing forecast
accuracy through parameter tuning.

REFERENCES
[1] Y. C. Lee, S.-H. Jung, and H.-H. Won, ‘‘WonDerM: Skin lesion classifica-
tion with fine-tuned neural networks,’’ 2018, arXiv:1808.03426.
[2] U. Jamil, M. U. Akram, S. Khalid, S. Abbas, and K. Saleem, ‘‘Computer
based melanocytic and nevus image enhancement and segmentation,’’
BioMed Res. Int., vol. 2016, pp. 1–13, Jan. 2016.
TABLE 7. Comparison between undersampling and oversampling. [3] A. Mahbod, G. Schaefer, C. Wang, R. Ecker, and I. Ellinge, ‘‘Skin lesion
classification using hybrid deep neural networks,’’ in Proc. IEEE Int. Conf.
Acoust., Speech Signal Process. (ICASSP), May 2019, pp. 1229–1233.
[4] K. Pai and A. Giridharan, ‘‘Convolutional neural networks for classify-
ing skin lesions,’’ in Proc. TENCON IEEE Region 10 Conf. (TENCON),
Oct. 2019, pp. 1794–1796.
[5] A. S. Shete, A. S. Rane, P. S. Gaikwad, and M. H. Patil, ‘‘Detection of skin
cancer using CNN algorithm,’’ Int. J., vol. 6, no. 5, pp. 1–4, 2021.
[6] M. Vidya and M. V. Karki, ‘‘Skin cancer detection using machine learning
Figure 19 shows the confusion matrix for undersam- techniques,’’ in Proc. IEEE Int. Conf. Electron., Comput. Commun. Tech-
nol. (CONECCT), Jul. 2020, pp. 1–5.
pling technique. The accuracy obtained by using under-
[7] H. Nahata and S. P. Singh, ‘‘Deep learning solutions for skin cancer detec-
sampling technique is 83% and f1 measure of 84%. tion and diagnosis,’’ in Machine Learning with Health Care Perspective.
Figure 20 shows the confusion matrix for oversampling Cham, Switzerland: Springer, 2020, pp. 159–182.
technique. [8] P. Wighton, T. K. Lee, H. Lui, D. I. McLean, and M. S. Atkins, ‘‘General-
izing common tasks in automated skin lesion diagnosis,’’ IEEE Trans. Inf.
The training and testing ratio were conducted on different Technol. Biomed., vol. 15, no. 4, pp. 622–629, Jul. 2011.
ranges like 80:20, 70:30 and 40:60. as shown in Table 5. [9] J. Saeed and S. Zeebaree, ‘‘Skin lesion classification based on deep con-
Table 6 shows the comparison between undersampling and volutional neural networks architectures,’’ J. Appl. Sci. Technol. Trends,
vol. 2, no. 1, pp. 41–51, Mar. 2021.
oversampling technique. [10] Y. Li, A. Esteva, B. Kuprel, R. Novoa, J. Ko, and S. Thrun, ‘‘Skin cancer
detection and tracking using data synthesis and deep learning,’’ 2016,
VI. COMPARISON BETWEEN EXISTING METHOD arXiv:1612.01074.
[11] V. Badrinarayanan, A. Kendall, and R. Cipolla, ‘‘SegNet: A deep convolu-
Our proposed work performs better when compare with other tional encoder–decoder architecture for image segmentation,’’ IEEE Trans.
CNN models that were recently published. We noticed that, Pattern Anal. Mach. Intell., vol. 39, no. 12, pp. 2481–2495, Dec. 2017.
the AUC score of 0.912 higher than the result in [17], 4% [12] P. Tschandl, C. Rosendahl, and H. Kittler, ‘‘The HAM10000 dataset, a
large collection of multi-source dermatoscopic images of common pig-
higher than the work carried out in [18] and 5% higher when
mented skin lesions,’’ Sci. Data, vol. 5, no. 1, pp. 1–9, Aug. 2018.
compared with [19]. [13] K. M. Hosny, M. A. Kassem, and M. M. Foaud, ‘‘Skin cancer classification
using deep learning and transfer learning,’’ in Proc. 9th Cairo Int. Biomed.
VII. CONCLUSION Eng. Conf. (CIBEC), Dec. 2018, pp. 90–93.
[14] A. Javaid, M. Sadiq, and F. Akram, ‘‘Skin cancer classification using image
Skin cancer is one of the illnesses that is spreading the processing and machine learning,’’ in Proc. Int. Bhurban Conf. Appl. Sci.
quickest on the earth. Skin cancer is mostly brought on by Technol. (IBCAST), Jan. 2021, pp. 439–444.

VOLUME 11, 2023 50213

 H. L. Gururaj et al.: DeepSkin: A Deep Learning Approach for Skin Cancer Classification

[15] R. Ashraf, I. Kiran, T. Mahmood, A. U. R. Butt, N. Razzaq, and Z. Farooq, A. NAGARJUN received the B.E. degree in
‘‘An efficient technique for skin cancer classification using deep learning,’’ computer science and engineering from the
in Proc. IEEE 23rd Int. Multitopic Conf. (INMIC), Nov. 2020, pp. 1–5. Sri Jayachamarajendra College of Engineering,
[16] M. Uckuner and H. Erol, ‘‘A new deep learning model for skin cancer clas- Mysuru, and the M.Tech. degree in data science
sification,’’ in Proc. 6th Int. Conf. Comput. Sci. Eng. (UBMK), Sep. 2021, from the JSS Science and Technology University,
pp. 27–31. Mysuru. He was a Project Intern with Vigyanlabs,
[17] Y. Filali, H. E. Khoukhi, M. A. Sabri, and A. Aarab, ‘‘Analysis and Mysuru. He is currently an Assistant Professor
classification of skin cancer based on deep learning approach,’’ in Proc. with the Department of Information Science and
Int. Conf. Intell. Syst. Comput. Vis. (ISCV), May 2022, pp. 1–6.
Engineering, JSS Science and Technology Uni-
[18] D. Yousra, A. B. Abdelhakim, and B. A. Mohamed, ‘‘Transfer learning
versity. He has published papers in international
for automated melanoma classification system: Data augmentation,’’ in
Proc. Int. Conf. Smart City Appl. Cham, Switzerland: Springer, Mar. 2023, conferences. His research interests include machine learning, artificial intel-
pp. 311–326. ligence, and image processing.
[19] T. Mazhar, I. Haq, A. Ditta, S. A. H. Mohsan, F. Rehman, I. Zafar,
J. A. Gansau, and L. P. W. Goh, ‘‘The role of machine learning and deep
learning approaches for the detection of skin cancer,’’ Healthcare, vol. 11,
no. 3, p. 415, Feb. 2023.

V. N. MANJUNATH ARADHYA received the

M.S. and Ph.D. degrees in computer science
from the University of Mysore, Mysuru, India,
H. L. GURURAJ (Senior Member, IEEE) received
in 2004 and 2007, respectively. He was a Post-
the Ph.D. degree in computer science and engi-
doctoral Research with the University of Genoa,
neering from Visvesvaraya Technological Univer-
Italy, from 2009 to 2010. He is currently a Pro-
sity, Belagavi, in 2019. He has 13 years of teaching
fessor with the Department of Computer Appli-
and research experience with UG and PG levels.
cations, JSS Science and Technology University,
He has published more than 75 research articles in
India. He is guiding four Ph.D. students and com-
peer-reviewed and reputed international journals.
pleted 12 Ph.D. students. He has published more
He has authored one book on network simulators
than 180 research papers in reputed international journals, conference pro-
and one edited book on blockchain by Springer
ceedings, and edited books. His research interests include pattern recog-
Publishers. He has presented more than 30 papers
nition, image processing, the Internet of Things (IoT), document image
at various international conferences. He is a Senior Member of ACM.
analysis, computer vision, machine intelligence, the applications of linear
He received a Young Scientist International Travel Support ITS-SERB,
algebra for the solution of engineering problems, the bi-clustering of gene
Department of Science and Technology, Government of India, in December
expression data, and web data analysis and understanding. He was a recipient
2016. He was a recipient of best paper awards at various national and
of the Young Indian Research Scientist Award from the Italian Ministry of
international conferences. He was appointed as an ACM Distinguish Speaker
Education, University and Research, Italy, in 2009. He also received the
(2018–2021) by the ACM U.S. Council, and he is one among 15 speakers
prestigious Award for Research Publications (ARP) from the Government
across India. He was a reviewer for various journals and conferences. He has
of Karnataka, in 2017.
delivered more than 100 technical talks across the globe. He has been hon-
ored as the session chair, a keynote speaker, a TPC member, and an advisory
committee member at national and international seminars, workshops, and
conferences across the globe.

FRANCESCO FLAMMINI (Senior Member,

IEEE) received the master’s (Cum Laude) and
N. MANJU received the B.E. degree in Ph.D. degrees (Hons.) in computer engineering
computer science and engineering from the from the University of Naples Federico II, Italy,
Sri Jayachamarajendra College of Engineering, in 2003 and 2006, respectively. He is currently a
Mysuru, the M.Tech. degree in computer network Professor of trustworthy autonomous systems at
engineering from NIE, Mysuru, and the Ph.D. the University of Applied Sciences and Arts of
degree from Visvesvaraya Technological Univer- Southern Switzerland, where he is affiliated with
sity. He is currently an Associate Professor with Dalle Molle Institute for Artificial Intelligence
the Department of Information Science and Engi- (IDSIA). He has worked for 15 years in industry
neering, JSS Science and Technology University, (Ansaldo STS, IPZS), in verification and validation, cyber-security, and
Mysuru. He has published papers in peer-reviewed innovation management. Since 2018, he has been working as an Associate
international journals and international conferences. His research interests Professor at Linnaeus University, Sweden. Since 2020, he has been a Full
include machine learning, computational intelligence, and computer net- Professor of computer science with a focus on cyber-physical systems at
works. He is a member of BOE and BOS, and a Life Member of ISTE. He was Mälardalen University, Sweden. He is an ACM Distinguished Speaker, an
a technical program committee member and the session chair at international IEEE Distinguished Visitor, a member-at-large of the IEE SMC BoG, and
conferences. He has organized many workshops and FDPs. He is a reviewer the Chair of the IEEE SMC TC on Homeland Security.
of international journals and international conferences.

50214 VOLUME 11, 2023