diagnostics

Article
An Efficient Deep Learning-Based Skin Cancer Classifier for an
Imbalanced Dataset
Talha Mahboob Alam 1,† , Kamran Shaukat 2,3, *,† , Waseem Ahmad Khan 4 , Ibrahim A. Hameed 5, * ,
Latifa Abd. Almuqren 6 , Muhammad Ahsan Raza 4 , Memoona Aslam 4 and Suhuai Luo 2

1 Department of Computer Science and Information Technology, Virtual University of Pakistan,

Lahore 54000, Pakistan
2 School of Information and Physical Sciences, The University of Newcastle, Newcastle, NSW 2308, Australia
3 Department of Data Science, University of the Punjab, Lahore 54890, Pakistan
4 School of Computer Science, National College of Business Administration & Economics,
Lahore 54660, Pakistan
5 Department of ICT and Natural Sciences, Norwegian University of Science and Technology,
7034 Trondheim, Norway
6 IS Department, College of Computer and Information Science, Princess Nourah Bint Abdulrahman University (PNU),
P.O. Box 84428, Riyadh 11671, Saudi Arabia
* Correspondence: [email protected] (K.S.); [email protected] (I.A.H.)
† These authors contributed equally to this work.

Abstract: Efficient skin cancer detection using images is a challenging task in the healthcare domain.
In today’s medical practices, skin cancer detection is a time-consuming procedure that may lead to a
patient’s death in later stages. The diagnosis of skin cancer at an earlier stage is crucial for the success
rate of complete cure. The efficient detection of skin cancer is a challenging task. Therefore, the
numbers of skilful dermatologists around the globe are not enough to deal with today’s healthcare.
Citation: Alam, T.M.; Shaukat, K.; The huge difference between data from various healthcare sector classes leads to data imbalance
Khan, W.A.; Hameed, I.A.; problems. Due to data imbalance issues, deep learning models are often trained on one class more than
Almuqren, L.A.; Raza, M.A.; others. This study proposes a novel deep learning-based skin cancer detector using an imbalanced
Aslam, M.; Luo, S. An Efficient Deep dataset. Data augmentation was used to balance various skin cancer classes to overcome the data
Learning-Based Skin Cancer imbalance. The Skin Cancer MNIST: HAM10000 dataset was employed, which consists of seven
Classifier for an Imbalanced Dataset. classes of skin lesions. Deep learning models are widely used in disease diagnosis through images.
Diagnostics 2022, 12, 2115. https://
Deep learning-based models (AlexNet, InceptionV3, and RegNetY-320) were employed to classify
doi.org/10.3390/diagnostics12092115
skin cancer. The proposed framework was also tuned with various combinations of hyperparameters.
Academic Editors: Dilbag Singh, The results show that RegNetY-320 outperformed InceptionV3 and AlexNet in terms of the accuracy,
Dinesh Kumar and Vijay Kumar F1-score, and receiver operating characteristic (ROC) curve both on the imbalanced and balanced
datasets. The performance of the proposed framework was better than that of conventional methods.
Received: 28 June 2022
Accepted: 29 August 2022
The accuracy, F1-score, and ROC curve value obtained with the proposed framework were 91%, 88.1%,
Published: 31 August 2022 and 0.95, which were significantly better than those of the state-of-the-art method, which achieved
85%, 69.3%, and 0.90, respectively. Our proposed framework may assist in disease identification,
Publisher’s Note: MDPI stays neutral
which could save lives, reduce unnecessary biopsies, and reduce costs for patients, dermatologists,
with regard to jurisdictional claims in
and healthcare professionals.
published maps and institutional affil-
iations.
Keywords: medical imaging; skin cancer; deep learning; disease diagnosis system; healthcare

Copyright: © 2022 by the authors.

Licensee MDPI, Basel, Switzerland. 1. Introduction
This article is an open access article The number of cancer patients is increasing due to smoking, environmental changes,
distributed under the terms and different types of radiation, viruses, alcohol, diet, and lifestyle [1]. The most common and
conditions of the Creative Commons hazardous type of cancer is skin cancer. Skin cancer can be in the form of unusual swelling
Attribution (CC BY) license (https://
of skin cells. Skin cancer is spreading worldwide and is a perilous disease [2]. The recorded
creativecommons.org/licenses/by/
new skin cancer case rate in the USA is around 5.4 million a year [3]. According to the
4.0/).

Diagnostics 2022, 12, 2115. https://doi.org/10.3390/diagnostics12092115 https://www.mdpi.com/journal/diagnostics

Diagnostics 2022, 12, 2115 2 of 16

WHO, annually diagnosed cases of melanoma have increased by 53%, and its mortality rate
will increase in the next decade. The failure of early diagnoses has shown a survival rate
of less than 14%, but detecting skin cancer at an early stage can increase the survival rate
to 97%, approximately [2,4,5]. The Skin Cancer Foundation reported that the skin cancer
problem is continuously growing. The most common type of cancer is non-melanocytic
such as basal cell carcinoma and squamous cell carcinoma.
Meanwhile, non-melanocytic skin cancer is the basic form to be found, such as basal
cell carcinoma and squamous cell carcinoma [6]. It is found that, in the United States,
around 1 million SCC cases and 4.3 million BCC cases are diagnosed each year, which
are still said to be underestimated. To improve the survival rate, early diagnoses are a
foundation, having a 99% correlation with overall survival, and survival is very poor once
the disease has penetrated far enough into the skin [7]. An early skin cancer diagnosis is
helpful for cancer treatment, although survival is poor once the disease progresses beyond
the skin. Medical specialists currently examine the infected person through visual means
with the help of polarised light magnification and dermoscopy [8]. The diagnoses also
depend on ethnicity, exposure to the sun, social habits, and the patient’s medical history.
Medical procedures for skin cancer diagnosis are time-consuming and very tough due
to the processing of biopsied lesions [2,9]. In the revolution of healthcare and medicine,
AI-enabled computer-aided diagnostics (CAD) solutions make significant contributions,
especially in medical imaging. Medical imaging is now a part of clinical practices such as
computed tomography (CT), magnetic resonance imaging (MRI), and ultrasound [10–12].
Dermoscopy, or less commonly confocal microscopy, plays a vital role in the in vivo vi-
sualisation of lesioned features, allowing for more accuracy in risk stratification in the
dermatological field. In various studies, AI-based algorithms exceed clinical performance
to detect disease in medical imaging [8,13,14].
Deep learning has recently provided end-to-end applications to identify breast cancer,
lung cancer, brain tumours, oesophagal cancer, and foot ulcer skin lesions. Imaging
techniques such as dermoscopy, CT, HRCT, and MRI have become helpful in diagnosing
cancer and are used to obtain data on skin cancer from patients worldwide [8,15]. Skin
imaging is a driving force behind skin lesion images and expert annotations for automated
CAD. High-speed internet, computing resources, and reliable cloud storage to maintain and
distribute skin cancer datasets have sparked the research interest in AI solutions for skin
cancer diagnosis [16,17]. These services can be applied to various computers, platforms,
and operating systems to transform them into cutting-edge medical devices. A skilful
dermatologist usually follows steps, from observation with the naked eye to dermoscopy
followed by a biopsy. However, this time-consuming procedure may lead patients to
advance to severe stages [18].
Training a classifier and learning the classes present in the dataset are much harder
when the dataset is imbalanced. This means that at least one of the classes present in
the dataset is significantly larger than the others. The problem is that the classes with
few instances have a low error cost and prior probability. Deep learning algorithms have
different learning strategies when trained on imbalanced datasets. Data augmentation
is a widely used method to overcome the data imbalance problem. Data augmentation
generates additional training data by transforming the input training data. Data augmen-
tation is especially beneficial for medical imaging [16,19]. Even for large datasets such as
ImageNet [20], it has been shown that data augmentation can be beneficial for very deep
architectures [16]. Additionally, data augmentation allows for an easy way to incorporate
prior knowledge about possible unseen data. Possible data augmentation schemes range
from simple additive or multiplicative image modifications such as intensity shifts to ge-
ometric transformations such as rotation, scaling, and elastic deformation, and synthetic
data generation [16,21].
Moreover, the main objective is to achieve accuracy in diagnosis, depending on the
clinician’s skills. Additionally, the best accuracy for diagnosing skin cancer is no more than
80%, as shown in Table 1. In addition to these difficulties, we do not have enough skilful
Diagnostics 2022, 12, 2115 3 of 16

dermatologists around the globe. A significant amount of work has been carried out to
rapidly develop image analysis algorithms to diagnose skin cancer at early stages and solve
the aforementioned problems. Most of these algorithms are parametric, i.e., they require
normalised data, but the data nature cannot be controlled. Hence, these methods cannot
diagnose the disease accurately.

Table 1. Characteristics of the CNNs’ architectures in the proposed framework.

Model Hidden Layers Image Size Parameters Learning Rate

AlexNet 8 227 × 227 200,132,679 [0.01, 0.001]
InceptionV3 48 299 × 299 22,126,759 [0.01, 0.001]
RegNetY-320 150 256 × 256 145,000,000 [0.01, 0.001]

Contributions: This research makes various contributions to skin cancer detection.

• The experiments were performed on the latest dataset. The Skin Cancer MNIST:
HAM10000 dataset presents cutting-edge images of the newest advancement in cancer
lesion detection. Previous studies employed smaller and noisier datasets that led to
less efficient results.
• Available skin cancer datasets are highly imbalanced, where multiple lesion cases
severely outnumber other lesion types. This paper presents an efficient and novel
deep learning-based skin cancer detector for handling imbalanced skin cancer de-
tection problems. Our results reveal that skin cancer detector performance was
significantly improved.
• Preprocessing, such as normalisation, image resizing, and data argumentation, was
conducted to eradicate the different biases in the dataset amid various classes.
• The performance of the proposed skin cancer detector was validated with state-of-the-
art detectors. The proposed skin cancer detector outperformed existing detectors. The
proposed skin cancer detector may assist in disease identification, which could save
lives, reduce unnecessary biopsies, and reduce costs for patients, dermatologists, and
healthcare professionals.
• The proposed deep learning-based skin cancer detector is high-performance, efficient,
time-efficient, and empowered with the latest advancement in deep learning and has
the least dependence on feature engineering.

2. Literature Review
Popescu et al. [22] presented a system based on the deep learning methodology and
collective intelligence. Various CNN-based models were employed on the HAM10000
dataset, which can differentiate skin lesions, including melanoma. They analysed the
various CNN models to maintain a weight matrix, and their elements were based on neural
network lesion classes. Furthermore, the accuracy of their system increased by about three
percent. Srinivasu et al. [23] proposed a deep learning-based model for analysing skin
disease detection by combining MobileNet and long short-term memory models. The
performance of the proposed hybrid model was also analysed to evaluate the growth of the
disease. Its results were compared with other state-of-the-art models such as fine-tuned
neural networks and CNNs. The proposed hybrid model achieved an accuracy of 85%
on the HAM10000 dataset. Khan et al. [24] presented a deep learning-based model for
effectively screening skin disease lesions. They performed the experiments using a mask
recurrent neural network (MASK-RNN), and a pyramid network was used with Resnet50
to extract and classify the SoftMax classifier. The proposed method exhibited efficient
performance on the HAM10000 dataset. In the study of Huang et al. [25], a lightweight
skin cancer detector was proposed to aid first-line medical care based on deep learning.
The HAM10000 dermoscopy dataset was employed for the training of the multiclass
classification model. Their proposed framework achieved an accuracy of 85.8%.
Diagnostics 2022, 12, 2115 4 of 16

Khan et al. [26] proposed a multiclass skin lesion classification method using local
colour-controlled histogram intensity values (LCcHIVs). Then, saliency was measured
using a novel deep saliency segmentation technique that includes a CNN, which consists
of ten layers. The heat map converts it into a binary image using the thresholding method.
They used an improved moth flame optimisation algorithm to avoid dimensionality to
select effective features. These features were used with multiple maximum correlation
analyses classified using a kernel extreme learning machine (KELM) classifier. The classifi-
cation performance was evaluated on the HAM10000 dataset and achieved an accuracy of
90.67%. Karl and Enrique [27] also presented a framework for skin cancer identification.
In their framework, transfer learning was applied to the convolutional neural network for
plain and hierarchical classification and used to differentiate between seven types of skin le-
sions. Xing et al. [28] presented a Categorical Relation-preserving Contrastive Knowledge
Distillation (CRCKD) that was used as a supervisor of the model. They presented a class-
guided contrastive distillation (CCD) module for closer image pairs from the same class
as a teacher while separating negative images from different classes. This showed higher
intra-class similarity and inter-class variance in teachers’ relational knowledge in a robust
and balanced manner. Extensive experiments on the HAM10000 dataset demonstrated the
superiority of the CRCKD method.
Saket et al. [29] presented a method for skin cancer identification. Their method
employed a better evaluation matrix technique than previous methodologies. They used the
MobileNet model for identifying cancer and HAM10000 employing transfer learning, and
their method achieved an accuracy of 83.1% for seven classes in the dataset. Ameri [30,31]
proposed a deep learning-based model for skin lesion classification. The proposed method
was trained on the HAM10000 dermoscopy image dataset to classify the melanoma and
non-melanoma lesions. Additionally, the deep CNN method was presented for image
classification. Transfer learning-based methods or deep learning-based models eliminate
the complex segmentation procedure of feature extraction. Andronescu et al. [32] developed
a model for identifying skin cancer using dermatoscopic images. A convolutional neural
network (CNN) detected images and patterns. The CNN works through three stages:
convolutional layer, pooling layer, and fully connected layer. HAM10000 was utilised,
containing 10,015 images, including seven skin lesions. These images were first resized to
90 × 120 pixels. Then, they were normalised. The dataset was divided into three parts:
training set, test set, and validation set. The CNN was used with a 3 × 3 kernel size and
one stride. A rectified linear unit (ReLU) was used as an activation function. Max pooling
with a size of 2 × 2 for each layer was used.

3. Methodology
First, the skin cancer dataset was obtained for a novel skin cancer detector and divided
into training and test sets. Further, augmentation techniques, i.e., rotating and flipping,
were applied to the training set to increase the data size to balance the classes. This training
dataset was shuffled well and augmented, i.e., reshaped and resized. This balanced dataset
was provided to the AlexNet, InceptionV3, and RegNetY-320 models for training. These
models were trained with 100% training accuracy. These were tested on the test dataset.
Their test accuracies were analysed, performing a comparison of their accuracies. The
proposed framework of our study is presented in Figure 1.

3.1. Dataset
Meaningful data are an essential component of deep learning. In this study, we
used the open-source Skin Cancer MNIST: HAM10000 dataset [33] consisting of 7 types
of skin lesions, namely: actinic keratoses and intraepithelial carcinoma/Bowen’s disease
(akiec), basal cell carcinoma (bcc), benign keratosis-like lesions (solar lentigines/seborrheic
keratoses and lichen planus-like keratoses (bkl)), dermatofibroma (df), melanoma (mel),
melanocytic nevi (nv), and vascular lesions (angiomas, angiokeratomas, pyogenic gran-
ulomas, and hemorrhage (vasc)). More than 50% of the lesions were confirmed through
 basal cell carcinoma (bcc), benign keratosis-like lesions (solar lentigines/seborrheic
keratoses and lichen planus-like keratoses (bkl)), dermatofibroma (df), melanoma (mel),
Diagnostics 2022, 12, 2115 melanocytic nevi (nv), and vascular lesions (angiomas, angiokeratomas, 5 pyogenic of 16

granulomas, and hemorrhage (vasc)). More than 50% of the lesions were confirmed
through histopathology (histo). The ground truth for the rest of the cases was either
histopathology
follow-up (histo).(follow_up),
examination The ground truth
expertforconsensus
the rest of(consensus),
the cases wasoreither follow-up by in
confirmation
examination (follow_up), expert
vivo confocal microscopy (confocal).consensus (consensus), or confirmation by in vivo confocal
microscopy (confocal).

Figure 1. Proposed framework.

Figure 1. Proposed framework.

3.2.3.2. Data
Data Balancing
Balancing
TheHAM10000
The HAM10000 dataset
dataset waswasemployed
employed in this study,
in this whichwhich
study, is proneis toprone
highlytoim- highly
balanced problems. Imbalanced data are a challenging problem while training a deep
imbalanced problems. Imbalanced data are a challenging problem while training a deep
learning model for a complex task [16,34]. Most deep learning models are designed to
learning model for a complex task [16,34]. Most deep learning models are designed to
work for classes with almost exact data for classification problems. When using a real-time
work for classes
dataset, withare
some events almost exact
rare, and wedata forhave
do not classification problems.
balanced data Whenespecially
for each class, using a real-
time dataset,
in the medicalsome events
domain [35].are
Thisrare, and we dataset
imbalanced do notoften
haveleads
balanced data or
to a biased forskewed
each class,
especially in the medical domain [35]. This imbalanced dataset often leads
prediction, affecting the model’s performance. Data augmentation can increase the sample to a biased or
skewed prediction,
size for affecting
those imbalanced the model’s
classes performance.
and produce a balancedData augmentation
dataset can increase
[16,21]. Predicting a
themodel
sampletrained
size with supervised
for those deep learning
imbalanced classesrelies
andonproduce
the diversity and the dataset
a balanced size of the[16,21].
dataset used
Predicting in training.
a model trained Thewith
relation between adeep
supervised rocket’s enginerelies
learning and the
onenormous amount
the diversity and the
of fuel used for a successful mission can represent the relation between the deep learning
size of the dataset used in training. The relation between a rocket’s engine and the
model and the data size used for training. Generally, deep learning models have many
enormous amount of fuel used for a successful mission can represent the relation between
hidden neurons for achieving high performance on complex tasks [36].
the deepThelearning
number ofmodel andparameters
trainable the data in
size used
a deep for training.
learning Generally,
model depends on the deep
numberlearning
models
of hidden neurons [37]. Hence, they need a large amount of data with huge diversity for [36].
have many hidden neurons for achieving high performance on complex tasks
 The number of trainable parameters in a deep learning model dep
Diagnostics 2022, 12, 2115
number of hidden neurons [37]. Hence, they need a large amount6 ofof16da
Diagnostics 2022, 12, 2115
diversity for training purposes [38,39]. Data augmentation has been used to
6 of 16
issues, i.e., increasing the training dataset’s size and diversity. For one cla
images,
The numberwhile of fortrainable
anotherparameters
class, it has in just
a deepa few hundred
learning modelimages.
depends Therefore,
on the
to
number insufficient
trainingofpurposes training
hidden [38,39].
neurons Dataofaugmentation
[37]. our model.
Hence, theyhasHence,
been aused
need weto used
large address
amount augmentation
these issues,
of data with techn
i.e., huge
increasing
diversity for the training
trainingto dataset’s
purposes size and
[38,39]. diversity. For
Data augmentation one class, it has 5000 images, while
image rotation
for another
balance
class, it hasthe
justtraining
our data,
a few hundred
as shown inhas
images.
been used
Figure 2. We to address
used data theseaug
issues, i.e., increasing dataset’s size Therefore, this may
and diversity. Forlead
onetoclass,
insufficient
it has 5000
increase
training of our dataset’s
our model. Hence, size
weby more
used than 30,000
augmentation and to
techniques make
such it balanced
as image rotation for ea
images, while for another class, it has just a few hundred images. Therefore, this may lead
to was done by randomly
of our cropping 256 we × 256 patches, flipping the images hor
to balance our data, as shown in Figure 2. We used data augmentation to increase our
insufficient training model. Hence, used augmentation
dataset’s size by more than 30,000 and to make it balanced for each class. This was done
techniques such as
rotating
image rotation them at different
to balance angles. We then obtained used more than 30,000to im
by randomly cropping 256our data,
× 256 as shown
patches, in Figure
flipping 2. We
the images data and
horizontally, augmentation
rotating
training
increase
them at our set
differentwith
dataset’s around
size
angles. by more
We then4000–5000
than 30,000
obtained images
more and
than for each
to30,000
make class.
it balanced
images Figure
for
for our each 3
training shows
set This th
class.
wasof done
with by randomly
classes
around before
4000–5000 cropping
and afterfor256
images the ×data
each 256 patches,
class. Figure flipping
augmentation.
3 shows the thedistribution
images horizontally,
of classes and
before and
rotating themafter
at the data augmentation.
different angles. We then obtained more than 30,000 images for our
training set with around 4000–5000 images for each class. Figure 3 shows the distribution
of classes before and after the data augmentation.

Figure 2. 2.
Figure Data augmentation
Data augmentation through imagerotation.
through image rotation.
Figure 2. Data augmentation through image rotation.

Figure 3. 3.
Figure Class distribution
Class distribution before andafter
before and afterdata
dataaugmentation.
augmentation.

3.3. AlexNet
3.3. AlexNet
The first CNN that became famous was AlexNet [40–42], which won the 2012 ILSVRC
Figure 3. Class
The first CNNdistribution before and
that became famous after data
was AlexNet augmentation.
[40–42], which won the 2012 ILSVRC
(ImageNet Large-Scale Visual Recognition Challenge), a prestigious challenge in the ma-
(ImageNet Large-Scale
chine learning Visual
field. It was Recognition
the first Challenge),
architecture that proveda the
prestigious challenge
power of CNNs in thein the
3.3.
machine AlexNet
context learning
of pattern recognition, becoming the state of the art in image classification, objectin the
field. It was the first architecture that proved the power of CNNs
context of pattern
detection, recognition,
object recognition, becoming
and human pose theestimation.
state of the art in image
AlexNet classification,
has eight weight layers,object
The first CNN that became famous was AlexNet [40–42], which won the
detection, object recognition, and human pose estimation. AlexNet has eight weight
(ImageNet
layers, Large-Scale
five convolutional Visual
layers (where Recognition
the ReLU per Challenge),
unit followsa theprestigious
convolution cha
machineand
operation), learning field.
three fully It was the
connected first
layers. architecture
The that layer
last is a SoftMax provedthatthe power
returns the o
context of
probability of pattern
belonging recognition,
to a certainbecoming
image class. theThis
state
is of
an the art in image
innovative ordering classif
of
Diagnostics 2022, 12, 2115 7 of 16

five convolutional layers (where the ReLU per unit follows the convolution operation),
and three fully connected layers. The last is a SoftMax layer that returns the probability of
belonging to a certain image class. This is an innovative ordering of operations, as in the
previous famous network, LeNet, a convolution was always followed by the non-linearity
and pooling, not by another convolution. The network has two parallel pipelines executed
in different GPUs to speed up the process. It is observed that the first convolution layer
uses a filter with a receptive field of 11 × 11, with stride 4 (number of pixels the filter shifts
from left to right and from up to down), immediately reducing the image spatially. The
receptive field diminishes to go deeper into the network to 5 × 5 and finally 3 × 3. This
means that the network initially tries to capture statistics for each pixel in a wider region.
As the filter size decreases, the image is down-sampled by max pooling operations, whereas
the number of filters increases from 96 to 256 and then 384. Thus, the data are compressed
spatially and up-sampled in depth. The model has many weights and memory needed for
keeping the feature maps during the forward/backward passes. The convolutional part of
the network requires more memory but less computation. The fully connected layers have
millions of weights, being the most computationally intensive part of the flow.
Two more novel properties are deployed in AlexNet: the ReLU activation instead
of tanh, and the local response normalisation. AlexNet empirically shows that training
with non-saturating non-linearity is faster and reaches a better convergence point. ReLUs
do not necessarily need input normalisation since, for learning to happen, it is enough
that some training examples have a positive input. However, using local normalisation
helps generalisation. The normalised response is defined for a unit obtained by applying
the filter to the position, defining the window size used for normalisation. Lastly, the
network is robust to some transformations by exposing it to an augmented dataset (flipped,
translated, reflected images where the label is preserved) and addressing overfitting by
applying dropout in the fully connected layers. Being the first successful deep network, the
representation properties of AlexNet have been studied extensively. There was already an
understanding that invariance and abstraction of features are created as we move deeper in
a network; the first layers in a convolutional network represent Gabor features. The higher
ones correspond to complex concepts in the image.

3.4. InceptionV3
InceptionV3 [43] is an updated version of GoogleNet [44], also called InceptionV1,
which reduces the number of parameters concerning state-of-the-art models 12 times.
The first version of the Inception architecture was introduced as GoogleNet in 2015. The
Inception module applies to different convolutions and max pooling to the same input
simultaneously to obtain multi-level features and combines them at the end of the module.
To compute them, GoogleNet uses three different filters of sizes 1 × 1, 3 × 3, and 5 × 5.
Furthermore, filter blocks were introduced to reduce dimensionality. It has also been
noticed that there was a problem of internal covariance shift, which means that when data
flow through the network, weights and parameters change data values, which could result
in being too big or too small. Sergey et al. [43] introduced batch normalisation, which
normalises data after each batch to overcome this problem. This new version of GoogleNet
is called InceptionV2. To scale the network, the 5 × 5 convolutional layer was factorised
into two consecutive 3 × 3 convolutional layers, and a new version of the network called
InceptionV3 was created.
Moreover, the architecture was re-factored to add factorisation convolution, modify the
auxiliary classifier, and introduce an efficient grid size reduction and the InceptionV3 ver-
sion. The factorisation convolution reduces the number of parameters without decreasing
the network efficiency. The factorisation techniques used in InceptionV3 are as follows.
Factorisation into smaller convolutions: This technique increases the number of
convolutional layers by stacking them to reduce the kernel size for each layer. For example,
one layer with a 7 × 7 kernel filter dimension has 49 parameters, while three layers
with 3 × 3 have 27. The number of parameters is reduced by 45%. With the usage of
 this technique, it is possible to modify a single Inception module (basic structure of
InceptionVX architectures) and reduce the number of network parameters.
Diagnostics 2022, 12, 2115 Factorisation into asymmetric convolutions: This technique reduces 8the of 16number
parameters using asymmetrical convolutional layers. The main concept is replacing
NxN filter with two consecutive layers of sizes 1 × N and N × 1, usually greater than 2
this technique,
For example, one itlayer
is possible
with to a 7modify a single
× 7 kernel Inception
filter dimensionmodulehas(basic
49 structure
parameters, of thewhile t
InceptionVX architectures) and reduce the number of network parameters.
layers with 1 × 7 and 7 × 1 have 14 parameters. The number of parameters is reduced
Factorisation into asymmetric convolutions: This technique reduces the number of
72%.parameters
With the usage of this technique,
using asymmetrical it is possible
convolutional layers. Theto main
modify a single
concept Inception
is replacing an mod
and reduce
NxN filterthe number
with of network
two consecutive layersparameters.
of sizes 1 × NThe and N auxiliary classifier,
× 1, usually greater than already
2N. pres
For example, one layer with a 7 × 7 kernel filter dimension has 49
since InceptionV1, had some modifications in InceptionV3. The V1 version has 2 auxiliaparameters, while two
layers with 1 × 7 and 7 × 1 have 14 parameters. The number of parameters is reduced
classifiers, while the V3 version has only 1 auxiliary classifier on top of the last 17 ×
by 72%. With the usage of this technique, it is possible to modify a single Inception
layers. The purpose
module and reduce ofthe
thenumber
auxiliary classifier
of network is also different:
parameters. firstly,
The auxiliary it is used
classifier, already to allow
a deeper network;
present with the V3
since InceptionV1, hadversion, it is used to
some modifications in regularise
InceptionV3.the The network.
V1 version Usually,
has am
2 auxiliary classifiers, while the V3 version has only 1 auxiliary
pooling layer is added to reduce the number of weights. Sometimes, this layer is nclassifier on top of the last
17 × if
efficient 17inserted
layers. The purpose
before of the auxiliary classifier
a convolutional layer, oris italso
is different:
too expensivefirstly, itifisinserted
used afte
to allow for a deeper network; with the V3 version, it is used to regularise the network.
convolutional layer. The efficient grid size reduction technique reduces these problems
Usually, a max pooling layer is added to reduce the number of weights. Sometimes,
creates
thisalayer
hybrid situation.
is not efficient ifEach layer
inserted concatenates
before a convolutionala convolutional
layer, or it is toolayer and aifpart of
expensive
max inserted
poolingafter
layer.
a convolutional layer. The efficient grid size reduction technique reduces
these problems. It creates a hybrid situation. Each layer concatenates a convolutional layer
and a part of the max pooling layer.
3.5. RegNetY-320
3.5. RegNetY-320
ResNet and its different versions have performed brilliantly in various compu
vision tasks. ResNet
ResNet and itswas a game-changer
different because itbrilliantly
versions have performed allowedinus to train
various extraordinar
computer
vision tasks. ResNet was a game-changer because it allowed us to train
deep neural networks with more than 150 layers effectively. Figure 4 depicts extraordinarily
deep neural networks with more than 150 layers effectively. Figure 4 depicts the bottleneck
bottleneck RegNet module based on the bottleneck ResNet building block proposed
RegNet module based on the bottleneck ResNet building block proposed to handle a
handle
large-scale image. image.
a large-scale

Figure 4. RegNetY-320 architecture.

Figure 4. RegNetY-320 architecture.

4. Results
4. Results
The retraining of the deep learning models was performed on an Intel i5 3.0 GHz. The
The retraining
framework chosen of
for the deepwas
this work learning models
TensorFlow, a deepwas performed
learning on aninIntel
library written Python i5 3.0 GH
The framework
and developed by Google. When performing the first stage of training, only original images written
chosen for this work was TensorFlow, a deep learning library
Pythonwereandused.developed by Google.
The oversampled When
images were performing
added theinfirst
to the dataset stage stage.
the second of training,
In o
addition to the training images, approximately 3000 (adjusted as
original images were used. The oversampled images were added to the dataset in a percentage of the total
input images) were used as the test set, regardless of the training set size. The test set was
second stage. In addition to the training images, approximately 3000 (adjusted a
only used at the end of each training session to evaluate the final accuracy of the network.
percentage of the
All images, bothtotal inputand
for training images)
testing, were used as sampled
were randomly the testfrom
set,the
regardless
dataset. Mostof the
of train
set size. The test set was
the hyperparameters wereonly
set used atdefault
to their the end of each
values. training was
The exception session to evaluate
the learning rate. the fi
accuracy of the network. All images, both for training and testing, were random
sampled from the dataset. Most of the hyperparameters were set to their default valu
The exception was the learning rate. The learning rate is probably the most import
 testing is not an option). When fine-tuning a network, the learning rate should be
decreased. Hence, it was changed from the default of 0.01 to 0.001.
Diagnostics 2022,The data from each class were split into test and training sets. The weightage for the9 of 16
12, 2115
test and training sets was almost 30% and 70% for balanced and imbalanced datasets. The
images were resized for each model. Training images were rescaled to 1/255 with a batch
The learning rate
size of 100 images. HAM10000 hasisvarious
probablyskin
the most
cancerimportant
imageshyperparameter
of imbalanced to classes
change ifwith
there is a
time constraint (i.e., when exhaustive parameter testing is not an option). When fine-tuning
10,000 images, including seven types of skin lesions. The first experiment employed the
a network, the learning rate should be decreased. Hence, it was changed from the default
AlexNet, InceptionV3, and
of 0.01 RegNetY-320 models on imbalanced data. The characteristics
to 0.001.
of the CNNs’ architectures employed
The data from each inclass
the were
proposed framework
split into are presented
test and training in Table
sets. The weightage for the
1. test and training sets was almost 30% and 70% for balanced and imbalanced datasets. The
The models were images were resized
trained on thefordataeachofmodel.
7000 Training
images images were rescaled
and tested on 3000toimages.
1/255 withThe a batch
size of 100 images. HAM10000 has various skin cancer images of imbalanced classes with
number of epochs was no more than 20, with a batch size of 100. We further trained the
10,000 images, including seven types of skin lesions. The first experiment employed the
models by steepening theInceptionV3,
AlexNet, learning rate. The AlexNet,
and RegNetY-320 modelsInceptionV3,
on imbalancedand data. RegNetY-320
The characteristics of
models were trained thewith
CNNs’ a architectures
learning rate of 0.01inand
employed achievedframework
the proposed an accuracy of 76%,in69%,
are presented Table 1.
and 80%, respectively. These models were also trained with a learning rate of 0.001 and The
The models were trained on the data of 7000 images and tested on 3000 images.
number of epochs was no more than 20, with a batch size of 100. We further trained the
achieved an accuracy of 76%, 77%, and 85%, respectively. Furthermore, the AlexNet,
models by steepening the learning rate. The AlexNet, InceptionV3, and RegNetY-320
InceptionV3, and RegNetY-320
models were trainedmodels
with awere
learningtrained with
rate of 0.01 andaachieved
learning rate of of0.01
an accuracy 76%,and
69%, and
achieved an F1-score 80%,ofrespectively.
52.2%, 49.9%, and 65.0%,
These models respectively.
were also trained with aThese
learningmodels were
rate of 0.001 andalso
achieved
trained with a learning an accuracy
rate ofof0.001
76%, and
77%, achieved
and 85%, respectively.
an F1-score Furthermore, the AlexNet,
of 60.2%, 63.7%, InceptionV3,
and 69.3%,
respectively. The and RegNetY-320
results show models
that the were performance
trained with a learning rate of 0.01 and achieved
of RegNetY-320 an F1-score
significantly
of 52.2%, 49.9%, and 65.0%, respectively. These models were also trained with a learning
increased when therate learning rate was changed. The complete results on the imbalanced
of 0.001 and achieved an F1-score of 60.2%, 63.7%, and 69.3%, respectively. The results
dataset are presented showin that
Figure 5.
the performance of RegNetY-320 significantly increased when the learning rate
was changed. The complete results on the imbalanced dataset are presented in Figure 5.

Figure 5. The comparison between

Figure 5. the AlexNet,
The comparison betweenInceptionV3, and RegNetY-320
the AlexNet, InceptionV3, modelsmodels
and RegNetY-320 on theon the
imbalanced dataset. imbalanced dataset.

The results obtained using the imbalanced dataset are not efficient. Therefore, a second
The results obtained using the imbalanced dataset are not efficient. Therefore, a
experiment was performed by employing image augmentation to obtain better results. The
second experimentvarious
was performed byofemploying
configurations image augmentation
the image augmentation to obtain
method are presented better
in Table 2.
results. The various configurations of the image augmentation method are presented in
Table 2.

Table 2. Various configurations of the image augmentation method.

Technique Configuration
Rotation (Random) [0°, 360°]
Translation (Random) [−10, 10] pixels
 Diagnostics 2022, 12, 2115 10 of 16
2115 10 of 16

Table 2. Various configurations of the image augmentation method.

Flipping Technique left to right Configuration
Shearing (Random) Rotation (Random) [−20°, 20°] [0◦ , 360◦ ]
Stretching (Random) Translation (Random) [1/1.3, 1.3] [−10, 10] pixels
Rescaling (Random) [1/1.6, 1.6]
Flipping left to right
The size of the images was increased
Shearing [−20◦ , 20
to 32,000 from 10,000 when image
(Random) ◦]
augmentation
Stretching (Random) [1/1.3, 1.3]
was applied. The models were trained on 22,000 images and tested on 10,000 images. As
the dataset was increased, the models could be trained better. The AlexNet, InceptionV3,
The size of the images was increased to 32,000 from 10,000 when image augmentation
and RegNetY-320 models were trained with a learning rate of 0.01 and achieved an
was applied. The models were trained on 22,000 images and tested on 10,000 images. As
accuracy of 76%, the78%, andwas
dataset 86%, respectively.
increased, the modelsThese
could models
be trainedwere
better.also trained InceptionV3,
The AlexNet, with a
learning rate of 0.001 and achieved
and RegNetY-320 modelsan accuracy
were of a76%,
trained with 85%,
learning rate and
of 0.0191%, respectively.
and achieved an accuracy
Furthermore, the AlexNet, InceptionV3, and RegNetY-320 models were trained with arate
of 76%, 78%, and 86%, respectively. These models were also trained with a learning
of 0.001
learning rate of 0.01 and and achieved
achieved ananF1-score
accuracy of
of76%, 85%,72.0%,
68.5%, and 91%,andrespectively. Furthermore, the
78.3%, respectively.
AlexNet, InceptionV3, and RegNetY-320 models were trained with a learning rate of 0.01
These models were also trained with a learning rate of 0.001 and achieved an F1-score of
and achieved an F1-score of 68.5%, 72.0%, and 78.3%, respectively. These models were also
60.2%, 77.1%, and trained
88.1%,with
respectively. Theofresults
a learning rate show
0.001 and that an
achieved theF1-score
performance
of 60.2%,of RegNetY-
77.1%, and 88.1%,
320 significantly increased
respectively.when the show
The results learning rate
that the was changed.
performance The complete
of RegNetY-320 results
significantly increased
when the learning rate was changed. The
obtained using the proposed framework are presented in Figure 6. complete results obtained using the proposed
framework are presented in Figure 6.

Figure 6. The performance

Figure 6.of the
The proposedofframework
performance the proposedbased on the
framework accuracy.
based on the accuracy.

It can be observed that the results of our proposed framework outperformed the
It can be observed that the results of our proposed framework outperformed the
state-of-the-art methods, as shown in Table 3. We employed a data augmentation technique
state-of-the-art methods,
to balanceas the shown
dataset ininourTable 3. We
proposed employed
framework. Neuralanetwork-based
data augmentation
architectures
technique to balance the dataset
are trained in on
much better ourbalanced
proposed framework.
data than imbalancedNeural network-based
data. However, we cannot find
balancedmuch
architectures are trained data in better
the real on
world, so we balanced
balanced the data
data than using data augmentation.
imbalanced data. However,Previous
studies claimed that clear convergence is expected to be revealed
we cannot find balanced data in the real world, so we balanced the data using data when training a classifier
increases the input data, while Table 3 supports our claim that there is a clear difference be-
augmentation. Previous studies claimed that clear convergence is expected to be revealed
tween the balanced and imbalanced dataset results. The results of our proposed framework
when training a classifier increases
also outperformed thethe inputstudies,
previous data, while Table
as shown 3 supports
in Table 1. our claim that
there is a clear difference between the balanced and imbalanced dataset results. The
results of our proposed framework also outperformed the previous studies, as shown in
Table 1.
Diagnostics 2022, 12, 2115 11 of 16

Table 3. The comparison of the performance on imbalanced data and the proposed framework.

Proposed Framework
Imbalanced Dataset
Model Learning Rate (Balanced Dataset)
Accuracy F1-Score ROC Value Accuracy F1-Score ROC Value
AlexNet 0.01 76% 52.2% 0.83 76% 68.5% 0.83
InceptionV3 0.01 69% 49.9% 0.75 78% 72.0% 0.84
RegNetY-320 0.01 80% 65.0% 0.85 86% 78.3% 0.92
AlexNet 0.001 76% 60.2% 0.83 76% 60.2% 0.83
InceptionV3 0.001 77% 63.7% 0.84 85% 77.1% 0.89
RegNetY-320 0.001 85% 69.3% 0.90 91% 88.1% 0.95

When analysing the problems with different algorithms, we often need to compare the
efficiency of each algorithm to determine which to choose. The ROC curve represents the
false positive rate (FPR) and true positive rate (TPR) under different threshold settings. Each
graph point represents T and FPR under a specific probability threshold. The threshold
ranges from 0 to 1. This is because FPR ranges from 0 to 1, as is obvious from its formula.
The ROC curve lies on (0,0) and (1,1) regardless of which model it is. The ideal TPR is 1,
which means a specific threshold exists where all positives are labelled as positives. The
ideal FPR is 0, which means a specific threshold exists where none of the negatives are
labelled as positives. Thus, (0,1) is the ideal point.
The advantage of the ROC curve is that it considers the balance of positive and negative
observations. TPR focuses on positive cases, and FPR focuses on negative cases. Therefore,
the ROC curve is a more balanced evaluation method. TPR and FPR, the two indicators
in the ROC curve, do not depend on a specific category distribution. Therefore, the ROC
curve has an outstanding feature compared with other evaluation methods. When the rate
of positive and negative observations in the test dataset changes, the ROC curve can remain
unchanged. In actual datasets, class imbalance often occurs. There are many more negative
observations than positive observations, and vice versa. The distribution of positive and
negative observations in the test dataset may also change. The ROC curve can show good
stability in this situation. The ROC curve was evaluated both on imbalanced data and the
proposed framework. In the case of an imbalanced dataset, the AlexNet, InceptionV3, and
RegNetY-320 models were trained with a learning rate of 0.01 and achieved an ROC curve
value of 0.83, 0.75, and 0.85, respectively. These models were also trained with a learning
rate of 0.001 and achieved an ROC curve value of 0.83, 0.84, and 0.90, respectively.
In contrast, using the proposed framework, the AlexNet, InceptionV3, and RegNetY-
320 models were trained with a learning rate of 0.01 and achieved ROC curve values of
0.83, 0.84, and 0.92, respectively. These models were also trained with a learning rate of
0.001 and achieved an ROC curve value of 0.83, 0.89, and 0.95, respectively. The accuracy
of the models concerning each class is presented in Table 4. The results show that the
performance of the models significantly increased using the proposed framework-based
ROC curve. The complete results obtained using the proposed framework based on the
ROC curve are presented in Figure 7.

Table 4. The accuracy of the models for each class.

Average
Model akiec bcc bkl df mel nv vasc
Accuracy
AlexNet 57.9 76.2 70.1 67.3 68.0 94.9 98.0 76.0%
InceptionV3 75.6 82.2 80.3 82.4 80.5 95.0 99.0 85.0%
RegNetY-
80.7 84.9 88.5 89.4 94.5 99.0 100.0 91.0%
320
Diagnostics 2022, 12, 2115 12 of 16
Diagnostics 2022, 12, 2115 12 of 16

Figure 7. (A) ROC curve obtained using AlexNet, InceptionV3, and RegNetY-320 with a learning
Figure 7. (A) ROC curve obtained using AlexNet, InceptionV3, and RegNetY-320 with a learning rate
rate of 0.01; (B) ROC curve obtained using AlexNet, InceptionV3, and RegNetY-320 with a learning
of 0.01; (B) ROC curve obtained using AlexNet, InceptionV3, and RegNetY-320 with a learning rate
rate of 0.001.
of 0.001.
5. Discussion
5. Discussion
The accuracy
The accuracyachieved
achievedon onthetheHAM10000
HAM10000 imbalanced
imbalanced dataset
dataset withwith
RegNetY-320
RegNetY-320 was
85%, while the performance improved to 91% after
was 85%, while the performance improved to 91% after the proposed framework was the proposed framework was
employed. Because the size of images also increased from 10,000
employed. Because the size of images also increased from 10,000 to 32,000 images, it to 32,000 images, it was
also also
was concluded
concluded thatthat
the the
performance
performance cancanbebeincreased
increasedbybyincreasing
increasingthe thedataset
dataset size.size.
Furthermore, neural network-based architectures performed
Furthermore, neural network-based architectures performed better on a balanced datasetbetter on a balanced dataset
for classification
for classification problems.
problems. Hence,Hence, the the performance
performance of of models
models is is directly
directly proportional
proportional to to
the size
the size of of the
the dataset.
dataset. The
The results
results obtained using ResNet
obtained using ResNet are are better
better thanthan those
those ofof AlexNet
AlexNet
InceptionV3. The
and InceptionV3. Thenumber
numberofoftrainable
trainableparameters
parametersininAlexNet
AlexNet is is 200,132,679,
200,132,679, leading
leading to
to an accuracy of 76%. This adds more evidence to the accuracy
an accuracy of 76%. This adds more evidence to the accuracy of the number of trainable of the number of trainable
parameters in the the neural
neural network.
network. However, when we trained InceptionV3 with just
22,126,759 trainable
22,126,759 trainableparameters,
parameters,we weshowed
showedananunexpected
unexpected behaviour
behaviour with
with anan accuracy
accuracy of
of 78%.
78%. This This exception
exception showsshowsthat that the accuracy
the accuracy dependsdepends
on theon the number
number of parameters.
of parameters. Still, it
Still,
is more it isdependent
more dependent on the architecture
on the architecture of the network,
of the network, i.e., the sequence
i.e., the sequence of layers,ofnumber
layers,
number
of of convolutional
convolutional layers,of
layers, number number
connectedof connected
layers, and layers, and the
the pattern theypattern they are
are connected
connected
in. When the in. learning
When therate learning rate of RegNetY-320
of RegNetY-320 was changed wasfromchanged0.01 from 0.01its
to 0.001, to accuracy
0.001, its
increased from 86%from
accuracy increased to 91% 86% in to
2091%
epochs
in 20with a batch
epochs withsize of 100.
a batch size This shows
of 100. that
This it might
shows that
be evident
it might bethat the accuracy
evident that the increases
accuracy with a decreasing
increases learning rate,
with a decreasing or there
learning is still
rate, vacant
or there is
space in thisspace
still vacant network fornetwork
in this more learningfor moreandlearning
better accuracy.
and better When we changed
accuracy. When we the changed
learning
rate
the of AlexNet
learning from
rate of0.01 to 0.001,
AlexNet its accuracy
from improved
0.01 to 0.001, by fractions,
its accuracy showingby
improved thatfractions,
a model
with
showing a slower
thatlearning
a model rate can extract
with a slower more featuresrate
learning andcaninformation
extract more from the dataset.
features and
The performance
information from the dataset. shown by AlexNet, InceptionV3, and RegNetY-320 after training on
the imbalanced dataset was not better than that of the proposed
The performance shown by AlexNet, InceptionV3, and RegNetY-320 after training framework, even at the
same
on thelearning
imbalancedrate ofdataset
0.001, with
was notan epoch
bettersize
thanofthat
20 and a batch
of the size offramework,
proposed 100. The accuracies
even at
the same learning rate of 0.001, with an epoch size of 20 and a batch size of 100. The
Diagnostics 2022, 12, 2115 13 of 16

of AlexNet, InceptionV3, and RegNetY-320 after utilising the proposed framework were
76%, 85%, and 91%, respectively, but decreased to 76%, 77%, and 85% after training on the
imbalanced dataset. Certain factors involve a significant decrease in the performance of
models. One of those reasons is that the dataset generated using the proposed framework
is much larger than the imbalanced dataset. The model can extract more features from a
larger dataset than it could with a smaller dataset. Secondly, larger data have more than
the model can learn, which is not the case with a smaller dataset. Due to skewed datasets
in a classification problem, the interest of the model builds higher towards classes with
more data and lower classes of a low data size. In classification problems, the model has to
draw boundaries between classes. If the model does not have enough data to differentiate
between classes, it starts confusing class boundaries, decreasing its performance [16,21,45].
A comparison of previous studies on the HAM10000 dataset is presented in Table 5.
The results also show that the deep learning-based models performed better on a
balanced dataset than on an imbalanced dataset. This might be due to the neural network’s
convolutional layers, weight updates, and deep learning. As the neural network does not
need pre-extracted features to be fed to the machine learning algorithm but extracts its
features based on exciting aspects of the class in the images, it might extract features that are
performing well in the dataset, making it more flexible, instead of extracting features that
perform well overall, which leads to overfitting [46–49]. It cannot be verified or falsified
whether the deep learning models were overfitted on this dataset, as the classifiers were
not tested on other datasets. The generalisability of the classifiers trained on this dataset is
unknown. The proposed framework should be generalised to similar tasks and datasets
of the same level of complexity. The demonstrated results depend on the dataset, which
indicates the biased behaviour of the proposed framework. The generalisation of the
proposed framework is indeed a limitation of our work.

Table 5. Comparison of previous studies on the HAM10000 dataset.

Reference Year Models Accuracy (%) F1-Score (%)

[50] 2022 DenseNet201 82.9 74.4%
[51] 2022 Wide-ShuffleNet 86.3 ____
[22] 2022 Collective Intelligence-based System 86.7 ____
[23] 2021 MobileNet V2-LSTM 90.7 ____
[24] 2021 Mask-RCNN 86.5 86.2
[25] 2021 EfficientNet-B4 85.8 ____
[26] 2021 Kernel extreme learning machine (KELM) classifier 90.6 ____
[27] 2021 DenseNet201 87.7 85.5
[28] 2021 CRCKD algorithm 85.6 76.4
[29] 2020 MobileNet 83.1 83.0
[30] 2020 Deep CNN 84.0 ____
[32] 2019 L2 regularisation 72.1 ____
Our study 2022 Proposed framework 91.0 88.1

6. Conclusions
Skin cancer is one of the deadliest diseases globally if not detected at the early stages.
Many deep learning-based applications using computer vision are designed to assist in
detecting skin cancer. This paper sought to find a solution for classifying skin lesions
using images with an efficient performance. A novel framework was proposed to solve
the problem of data imbalance. The classes in the dataset were not balanced, limiting
the performance of deep learning models. Data augmentation techniques are used to
increase the size of the dataset and resolve the data imbalance issue. Our proposed frame-
work was trained on the Skin Cancer MNIST: HAM10000 dataset. AlexNet, InceptionV3,
and RegNetY-320-based deep learning models were trained on balanced and imbalanced
datasets. The proposed framework was tuned on different hyperparameters, i.e., the learn-
ing rate, epochs, and batch size in which the learning rate was changed, but the epochs and
Diagnostics 2022, 12, 2115 14 of 16

batch size were fixed. The performance of the RegNetY-320 model was better than that of
AlexNet and InceptionV3 in terms of the accuracy and ROC curve both on the imbalanced
and balanced datasets.
Furthermore, the accuracy obtained using the proposed framework was 91%, which
was significantly better than the state-of-the-art method, which achieved 85%. In the
future, to see a convergence in the accuracy of RegNetY-320, it would be valuable to test
it on a larger training set. It would be interesting to compare the results of the proposed
framework with those of dermatologists for the clinical implementation of our proposed
framework in skin cancer identification. This would provide healthcare institutions with
guidance on when it is appropriate to use our proposed framework as a second opinion
or even replace the human factor. Furthermore, the proposed framework should also be
tested on other skin cancer datasets.

Author Contributions: Data curation, T.M.A., K.S., W.A.K., I.A.H. and L.A.A.; Formal analysis, K.S.,
W.A.K., I.A.H., L.A.A. and M.A.R.; Funding acquisition, I.A.H., K.S., L.A.A. and M.A.R.; Investigation,
T.M.A. and K.S.; Methodology, T.M.A., K.S., W.A.K., I.A.H. and L.A.A.; Project administration, K.S.,
M.A.R. and S.L.; Software, T.M.A., K.S., W.A.K., I.A.H. and L.A.A.; Supervision, K.S., W.A.K., I.A.H.,
L.A.A. and M.A.R.; Validation, K.S., W.A.K., I.A.H., L.A.A., M.A. and S.L.; Visualization, K.S., W.A.K.,
I.A.H., L.A.A., M.A.R., M.A. and S.L.; Writing—original draft, T.M.A., K.S., W.A.K., I.A.H., L.A.A.,
M.A.R., M.A. and S.L.; Writing—review & editing, T.M.A., K.S., W.A.K., I.A.H., L.A.A., M.A.R., M.A.
and S.L. All authors have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: The data will be provided upon reasonable request.
Acknowledgments: This research was funded by the Deanship of Scientific Research at Princess
Nourah bint Abdulrahman University through the Fast-track Research Funding Program.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Alam, T.M.; Khan, M.M.A.; Iqbal, M.A.; Abdul, W.; Mushtaq, M. Cervical cancer prediction through different screening methods
using data mining. IJACSA Int. J. Adv. Comput. Sci. Appl. 2019, 10, 388–396. [CrossRef]
2. Tavakolpour, S.; Daneshpazhooh, M.; Mahmoudi, H. Skin cancer: Genetics, immunology, treatments, and psychological care. In
Cancer Genetics and Psychotherapy; Springer: Berlin/Heidelberg, Germany, 2017; pp. 851–934.
3. Ferlay, J.; Colombet, M.; Soerjomataram, I.; Parkin, D.M.; Piñeros, M.; Znaor, A.; Bray, F. Cancer statistics for the year 2020: An
overview. Int. J. Cancer 2021, 149, 778–789. [CrossRef]
4. Brunssen, A.; Waldmann, A.; Eisemann, N.; Katalinic, A. Impact of skin cancer screening and secondary prevention campaigns
on skin cancer incidence and mortality: A systematic review. J. Am. Acad. Dermatol. 2017, 76, 129–139.e10. [CrossRef] [PubMed]
5. Niino, M.; Matsuda, T. Age-specific skin cancer incidence rate in the world. Jpn. J. Clin. Oncol. 2021, 51, 848–849. [CrossRef]
6. Cia˛żyńska, M.; Kamińska-Winciorek, G.; Lange, D.; Lewandowski, B.; Reich, A.; Sławińska, M.; Pabianek, M.; Szczepaniak, K.;
Hankiewicz, A.; Ułańska, M.; et al. The incidence and clinical analysis of non-melanoma skin cancer. Sci. Rep. 2021, 11, 4337.
[CrossRef] [PubMed]
7. Nikolouzakis, T.K.; Falzone, L.; Lasithiotakis, K.; Krüger-Krasagakis, S.; Kalogeraki, A.; Sifaki, M.; Spandidos, D.A.; Chrysos, E.;
Tsatsakis, A.; Tsiaoussis, J. Current and future trends in molecular biomarkers for diagnostic, prognostic, and predictive purposes
in non-melanoma skin cancer. J. Clin. Med. 2020, 9, 2868. [CrossRef] [PubMed]
8. Goyal, M.; Knackstedt, T.; Yan, S.; Hassanpour, S. Artificial intelligence-based image classification for diagnosis of skin cancer:
Challenges and opportunities. Comput. Biol. Med. 2020, 127, 104065. [CrossRef]
9. Chan, S.; Reddy, V.; Myers, B.; Thibodeaux, Q.; Brownstone, N.; Liao, W. Machine learning in dermatology: Current applications,
opportunities, and limitations. Dermatol. Ther. 2020, 10, 365–386. [CrossRef]
10. Kousis, I.; Perikos, I.; Hatzilygeroudis, I.; Virvou, M. Deep Learning Methods for Accurate Skin Cancer Recognition and Mobile
Application. Electronics 2022, 11, 1294. [CrossRef]
11. Nawaz, M.; Mehmood, Z.; Nazir, T.; Naqvi, R.A.; Rehman, A.; Iqbal, M.; Saba, T. Skin cancer detection from dermoscopic images
using deep learning and fuzzy k-means clustering. Microsc. Res. Tech. 2022, 85, 339–351. [CrossRef]
12. Bechelli, S.; Delhommelle, J. Machine Learning and Deep Learning Algorithms for Skin Cancer Classification from Dermoscopic
Images. Bioengineering 2022, 9, 97. [CrossRef] [PubMed]
Diagnostics 2022, 12, 2115 15 of 16

13. Phillips, M.; Greenhalgh, J.; Marsden, H.; Palamaras, I. Detection of malignant melanoma using artificial intelligence: An
observational study of diagnostic accuracy. Dermatol. Pract. Concept. 2020, 10, e2020011. [CrossRef] [PubMed]
14. Alfi, I.A.; Rahman, M.M.; Shorfuzzaman, M.; Nazir, A. A Non-Invasive Interpretable Diagnosis of Melanoma Skin Cancer Using
Deep Learning and Ensemble Stacking of Machine Learning Models. Diagnostics 2022, 12, 726. [CrossRef] [PubMed]
15. Reis, H.C.; Turk, V.; Khoshelham, K.; Kaya, S. InSiNet: A deep convolutional approach to skin cancer detection and segmentation.
Med. Biol. Eng. Comput. 2022, 60, 643–662. [CrossRef]
16. Mikołajczyk, A.; Grochowski, M. Data augmentation for improving deep learning in image classification problem. In Proceedings
of the 2018 International Interdisciplinary PhD Workshop (IIPhDW), Swinoujscie, Poland, 9–12 May 2018; pp. 117–122.
17. Santos, M.A.; Munoz, R.; Olivares, R.; Filho, P.P.R.; del Ser, J.; de Albuquerque, V.H.C. Online heart monitoring systems on the
internet of health things environments: A survey, a reference model and an outlook. Inf. Fusion 2020, 53, 222–239. [CrossRef]
18. Wolner, Z.J.; Yélamos, O.; Liopyris, K.; Rogers, T.; Marchetti, M.A.; Marghoob, A.A. Enhancing skin cancer diagnosis with
dermoscopy. Dermatol. Clin. 2017, 35, 417–437. [CrossRef]
19. Srinivas, C.; KS, N.P.; Zakariah, M.; Alothaibi, Y.A.; Shaukat, K.; Partibane, B.; Awal, H. Deep Transfer Learning Approaches in
Performance Analysis of Brain Tumor Classification Using MRI Images. J. Healthc. Eng. 2022, 2022, 3264367. [CrossRef]
20. Russakovsky, O.; Deng, J.; Su, H.; Krause, J.; Satheesh, S.; Ma, S.; Huang, Z.; Karpathy, A.; Khosla, A.; Bernstein, M.; et al.
Imagenet large scale visual recognition challenge. Int. J. Comput. Vis. 2015, 115, 211–252. [CrossRef]
21. Saini, M.; Susan, S. Deep transfer with minority data augmentation for imbalanced breast cancer dataset. Appl. Soft Comput. 2020,
97, 106759. [CrossRef]
22. Popescu, D.; El-khatib, M.; Ichim, L. Skin Lesion Classification Using Collective Intelligence of Multiple Neural Networks. Sensors
2022, 22, 4399. [CrossRef]
23. Srinivasu, P.N.; SivaSai, J.G.; Ijaz, M.F.; Bhoi, A.K.; Kim, W.; Kang, J.J. Classification of skin disease using deep learning neural
networks with MobileNet V2 and LSTM. Sensors 2021, 21, 2852. [CrossRef] [PubMed]
24. Khan, M.A.; Zhang, Y.-D.; Sharif, M.; Akram, T. Pixels to classes: Intelligent learning framework for multiclass skin lesion
localization and classification. Comput. Electr. Eng. 2021, 90, 106956. [CrossRef]
25. Huang, H.W.; Hsu, B.W.Y.; Lee, C.H.; Tseng, V.S. Development of a light-weight deep learning model for cloud applications and
remote diagnosis of skin cancers. J. Dermatol. 2021, 48, 310–316. [CrossRef] [PubMed]
26. Khan, M.A.; Sharif, M.; Akram, T.; Damaševičius, R.; Maskeliūnas, R. Skin lesion segmentation and multiclass classification using
deep learning features and improved moth flame optimization. Diagnostics 2021, 11, 811. [CrossRef] [PubMed]
27. Thurnhofer-Hemsi, K.; Domínguez, E. A convolutional neural network framework for accurate skin cancer detection. Neural
Process. Lett. 2021, 53, 3073–3093. [CrossRef]
28. Xing, X.; Hou, Y.; Li, H.; Yuan, Y.; Li, H.; Meng, M.Q.-H. Categorical Relation-Preserving Contrastive Knowledge Distillation for
Medical Image Classification. In International Conference on Medical Image Computing and Computer-Assisted Intervention; Springer:
Berlin/Heidelberg, Germany, 2021; pp. 163–173.
29. Chaturvedi, S.S.; Gupta, K.; Prasad, P.S. Skin lesion analyser: An efficient seven-way multi-class skin cancer classification using
MobileNet. In International Conference on Advanced Machine Learning Technologies and Applications; Springer: Berlin/Heidelberg,
Germany, 2020; pp. 165–176.
30. Ameri, A. A deep learning approach to skin cancer detection in dermoscopy images. J. Biomed. Phys. Eng. 2020, 10, 801. [CrossRef]
31. Khushi, M.; Shaukat, K.; Alam, T.M.; Hameed, I.A.; Uddin, S.; Luo, S.; Yang, X.; Reyes, M.C. A comparative performance analysis
of data resampling methods on imbalance medical data. IEEE Access 2021, 9, 109960–109975. [CrossRef]
32. Andronescu, A.; Nastac, D.; Tiplica, G. Skin Anomaly Detection Using Classification Algorithms. In Proceedings of the 2019
IEEE 25th International Symposium for Design and Technology in Electronic Packaging (SIITME), Cluj-Napoca, Romania,
23–26 October 2019; pp. 299–303.
33. Tschandl, P.; Rosendahl, C.; Kittler, H. The HAM10000 dataset, a large collection of multi-source dermatoscopic images of
common pigmented skin lesions. Sci. Data 2018, 5, 180161. [CrossRef]
34. Sabanci, K.; Aslan, M.F.; Ropelewska, E.; Unlersen, M.F. A convolutional neural network-based comparative study for pepper seed
classification: Analysis of selected deep features with support vector machine. J. Food Process Eng. 2021, 45, e13955. [CrossRef]
35. Johnson, J.M.; Khoshgoftaar, T.M. Survey on deep learning with class imbalance. J. Big Data 2019, 6, 27. [CrossRef]
36. Bhimavarapu, U.; Battineni, G. Skin Lesion Analysis for Melanoma Detection Using the Novel Deep Learning Model Fuzzy
GC-SCNN. Healthcare 2022, 10, 962. [CrossRef] [PubMed]
37. Codella, N.C.; Nguyen, Q.-B.; Pankanti, S.; Gutman, D.A.; Helba, B.; Halpern, A.C.; Smith, J.R. Deep learning ensembles for
melanoma recognition in dermoscopy images. IBM J. Res. Dev. 2017, 61, 5:1–5:15. [CrossRef]
38. Khan, A.; Sohail, A.; Zahoora, U.; Qureshi, A.S. A survey of the recent architectures of deep convolutional neural networks. Artif.
Intell. Rev. 2020, 53, 5455–5516. [CrossRef]
39. Batool, D.; Shahbaz, M.; Asif, H.S.; Shaukat, K.; Alam, T.M.; Hameed, I.A.; Ramzan, Z.; Waheed, A.; Aljuaid, H.; Luo, S. A Hybrid
Approach to Tea Crop Yield Prediction Using Simulation Models and Machine Learning. Plants 2022, 11, 1925. [CrossRef]
40. Krizhevsky, A.; Sutskever, I.; Hinton, G.E. Imagenet classification with deep convolutional neural networks. Adv. Neural Inf.
Process. Syst. 2012, 25, 1097–1105. [CrossRef]
41. Alam, T.M.; Shaukat, K.; Khelifi, A.; Khan, W.A.; Raza, H.M.E.; Idrees, M.; Luo, S.; Hameed, I.A. Disease diagnosis system using
IoT empowered with fuzzy inference system. Comput. Mater. Contin. 2022, 70, 5305–5319.
Diagnostics 2022, 12, 2115 16 of 16

42. Yang, X.; Khushi, M.; Shaukat, K. Biomarker CA125 Feature Engineering and Class Imbalance Learning Improves Ovarian Cancer
Prediction. In Proceedings of the 2020 IEEE Asia-Pacific Conference on Computer Science and Data Engineering (CSDE), Gold
Coast, Australia, 16–18 December 2020; pp. 1–6.
43. Szegedy, C.; Vanhoucke, V.; Ioffe, S.; Shlens, J.; Wojna, Z. Rethinking the inception architecture for computer vision. In Proceedings
of the IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, NV, USA, 27–30 June 2016; pp. 2818–2826.
44. Szegedy, C.; Liu, W.; Jia, Y.; Sermanet, P.; Reed, S.; Anguelov, D.; Erhan, D.; Vanhoucke, V.; Rabinovich, A. Going deeper
with convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Boston, MA, USA,
7–12 June 2015; pp. 1–9.
45. Ali, I.S.; Mohamed, M.F.; Mahdy, Y.B. Data augmentation for skin lesion using self-attention based progressive generative
adversarial network. arXiv 2019, arXiv:1910.11960.
46. Krois, J.; Mahmood, A. Generalizability of deep learning models for dental image analysis. Sci. Rep. 2021, 11, 6102. [CrossRef]
47. Shrestha, A.; Mahmood, A. Review of deep learning algorithms and architectures. IEEE Access 2019, 7, 53040–53065. [CrossRef]
48. Shaukat, K.; Alam, T.M.; Ahmed, M.; Luo, S.; Hameed, I.A.; Iqbal, M.S.; Li, J.; Iqbal, M.A. A Model to Enhance Governance
Issues through Opinion Extraction. In Proceedings of the 2020 11th IEEE Annual Information Technology Electronics and Mobile
Communication Conference (IEMCON), Vancouver, BC, Canada, 4–7 November 2020; pp. 0511–0516.
49. Shaukat, K.; Alam, T.M.; Hameed, I.A.; Khan, W.A.; Abbas, N.; Luo, S. A Review on Security Challenges in Internet of Things
(IoT). In Proceedings of the 2021 26th International Conference on Automation and Computing (ICAC), Portsmouth, UK,
2–4 September 2021; pp. 1–6. [CrossRef]
50. Fraiwan, M.; Faouri, E. On the Automatic Detection and Classification of Skin Cancer Using Deep Transfer Learning. Sensors
2022, 22, 4963. [CrossRef]
51. Hoang, L.; Lee, S.-H.; Lee, E.-J.; Kwon, K.-R. Multiclass Skin Lesion Classification Using a Novel Lightweight Deep Learning
Framework for Smart Healthcare. Appl. Sci. 2022, 12, 2677. [CrossRef]