cancers

Article
DSCC_Net: Multi-Classification Deep Learning Models for
Diagnosing of Skin Cancer Using Dermoscopic Images
Maryam Tahir 1,† , Ahmad Naeem 2,† , Hassaan Malik 1,2 , Jawad Tanveer 3 , Rizwan Ali Naqvi 4, *,†
and Seung-Won Lee 5, *

1 Department of Computer Science, National College of Business Administration & Economics Lahore,
Multan Sub Campus, Multan 60000, Pakistan
2 Department of Computer Science, University of Management and Technology, Lahore 54000, Pakistan
3 Department of Computer Science and Engineering, Sejong University, Seoul 05006, Republic of Korea
4 Department of Intelligent Mechatronics Engineering, Sejong University, Seoul 05006, Republic of Korea
5 School of Medicine, Sungkyunkwan University, Suwon 16419, Republic of Korea
* Correspondence: rizwanali@sejong.ac.kr (R.A.N.); swleemd@g.skku.edu (S.-W.L.)
† These authors contributed equally to this work.

Simple Summary: This paper proposes a deep learning-based skin cancer classification network
(DSCC_Net) that is based on a convolutional neural network (CNN) and implemented on three
publicly available benchmark datasets (ISIC 2020, HAM10000, and DermIS). The proposed DSCC_Net
obtained a 99.43% AUC, along with a 94.17% accuracy, a recall of 93.76%, a precision of 94.28%,
and an F1-score of 93.93% in categorizing the four distinct types of skin cancer diseases. The
accuracies of ResNet-152, Vgg-19, MobileNet, and Vgg-16, EfficientNet-B0, and Inception-V3 are
89.68%, 92.51%, 91.46%, 89.12%, 89.46%, and 91.82%, respectively. The results showed that the
proposed DSCC_Net model performs better as compared to baseline models, thus offering significant
support to dermatologists and health experts to diagnose skin cancer.

Abstract: Skin cancer is one of the most lethal kinds of human illness. In the present state of the
health care system, skin cancer identification is a time-consuming procedure and if it is not diagnosed
Citation: Tahir, M.; Naeem, A.; Malik,
initially then it can be threatening to human life. To attain a high prospect of complete recovery, early
H.; Tanveer, J.; Naqvi, R.A.; Lee, S.-W.
detection of skin cancer is crucial. In the last several years, the application of deep learning (DL)
DSCC_Net: Multi-Classification
algorithms for the detection of skin cancer has grown in popularity. Based on a DL model, this work
Deep Learning Models for
Diagnosing of Skin Cancer Using
intended to build a multi-classification technique for diagnosing skin cancers such as melanoma
Dermoscopic Images. Cancers 2023, (MEL), basal cell carcinoma (BCC), squamous cell carcinoma (SCC), and melanocytic nevi (MN).
15, 2179. https://doi.org/10.3390/ In this paper, we have proposed a novel model, a deep learning-based skin cancer classification
cancers15072179 network (DSCC_Net) that is based on a convolutional neural network (CNN), and evaluated it on
three publicly available benchmark datasets (i.e., ISIC 2020, HAM10000, and DermIS). For the skin
Academic Editors: Mario Mascalchi
cancer diagnosis, the classification performance of the proposed DSCC_Net model is compared with
and Stefano Diciotti
six baseline deep networks, including ResNet-152, Vgg-16, Vgg-19, Inception-V3, EfficientNet-B0, and
Received: 7 March 2023 MobileNet. In addition, we used SMOTE Tomek to handle the minority classes issue that exists in this
Revised: 4 April 2023 dataset. The proposed DSCC_Net obtained a 99.43% AUC, along with a 94.17%, accuracy, a recall of
Accepted: 4 April 2023
93.76%, a precision of 94.28%, and an F1-score of 93.93% in categorizing the four distinct types of skin
Published: 6 April 2023
cancer diseases. The rates of accuracy for ResNet-152, Vgg-19, MobileNet, Vgg-16, EfficientNet-B0,
and Inception-V3 are 89.32%, 91.68%, 92.51%, 91.12%, 89.46% and 91.82%, respectively. The results
showed that our proposed DSCC_Net model performs better as compared to baseline models, thus
Copyright: © 2023 by the authors. offering significant support to dermatologists and health experts to diagnose skin cancer.
Licensee MDPI, Basel, Switzerland.
This article is an open access article Keywords: skin cancer; melanoma; deep learning; transfer learning; CNN; dermoscopic images
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).

Cancers 2023, 15, 2179. https://doi.org/10.3390/cancers15072179 https://www.mdpi.com/journal/cancers

Cancers 2023, 15, 2179 2 of 28

1. Introduction
The largest organ in the body is the skin, which saves the body from infection, heat, and
UV light, but the serious threat to human life is cancer. The human body may harbor various
kinds of cancer, and skin cancer is one of the deadliest and rapidly growing tumors. One
in every three cancers diagnosed is skin cancer and, according to Skin Cancer Foundation
Statistics, one in every five Americans will develop skin cancer in their lifetime [1–4]. In the
USA, there are more than 3.5 million new cases that appear every year, and that number of
cases is continuously increasing [3].
Many skin cancers begin in the upper layer of the skin. Skin cancer occurs when
skin cells divide and expand in an uncontrolled way. New skin cells usually develop
when old ones die or are damaged. When this process does not work correctly, cells grow
quickly in an unordered way. This is why these cells are known as a tumor, which is in
the form of a group of tissue [5,6]. It is caused by several factors, such as drinking alcohol,
smoking, allergies, viruses, changing environments, and ultraviolet (UV) light exposure.
Furthermore, skin cancer can also appear due to abnormal swellings on the body.
There are four different types of skin cancer: melanoma (MEL), melanocytic nevi (MN),
basal cell carcinoma (BCC), and squamous cell carcinoma (SCC). The most dangerous
category of cancer is MEL, because it spreads quickly to other organs. It arrives from the
skin cells that are called melanocytes. On the skin, melanocytes create dark pigments, and
these are mostly black and brown, while some are red, purple, and pink [7]. A melanoma
cell frequently spreads to another organ, such as the brain, liver, or lungs [8,9]. Due to
melanoma cancer, 10,000 deaths occur annually in the United States [10]. If it is identified
early, then melanoma can be treated as soon as possible. It is not more common than other
kinds of skin cancer. Melanocytic nevi (MN) happen in a pigmented mole that varies in a
variety of skin tone colors. It mostly occurs throughout childhood and the early years of
adult life, because the number of moles on one’s body increases up until the 30 to 40 years
of age. Basal cell carcinoma (BCC) is the most common type of skin cancer. These are
round cells that are created in the lower portion of the epidermis and normally grow slowly.
Approximately all BCC develops on areas of the body that have a lot of sun exposure,
including the face, neck, head, ears, back, and shoulders. Rarely, this type of skin cancer
migrates to other body areas, and forms due to the abnormal and uncontrolled growth of
cells. It may occur as a small, flesh-colored, or white tumor that may bleed. Squamous cell
carcinoma (SCC) comprises flat cells found in the upper portion of the epidermis. These
cancer cells can arise when cells grow uncontrollably. It may occur as a hard red mark
or open sore that may bleed easily. Although this type of skin cancer is not normally
dangerous, SCC can be found in numerous areas because it is usually generated by sun
exposure. Additionally, it may also develop on skin that has already been burned or
harmed by chemicals.
Skin cancer detection is a challenging process, and there are many different ways in
which doctors can find skin cancer. An experienced dermatologist uses a sequence of steps
to make a diagnosis, beginning with the naked eye detection of abnormal tumors, followed
by dermoscopy, which uses a magnifying lens to conduct an in-depth analysis of lesion pat-
terns, and the final step is biopsy [11,12]. Before the development of dermoscopic pictures,
most skilled dermatologists had a rate of success of only 60 percent in diagnosing skin can-
cer, but dermoscopic images raised success rates to between 75 percent and 84 percent [13].
Additionally, correct identification is unique and largely dependent on the skills of the
clinician [14]. The manual diagnosis of skin disorders is extremely difficult and stressful
for the patient [15]. Computer-aided detection systems support health professionals to
evaluated data garnered from dermoscopy method in situations where there is a shortage
of professional availability or diagnostic expertise [16,17].
Skin cancer is a huge problem that needs to be investigated as soon as possible. The
majority of people do not visit their dermatologist on a regular basis, which causes a fatally
delayed diagnosis. The diagnosis is a manual process that takes a lot of time and money.
However, diagnosis improved due to machine learning, and this can be useful in various
Cancers 2023, 15, 2179 3 of 28

ways. Skin cancer classification has been worked out using machine learning techniques,
such as the support vector machine (SVM) [18], the Naïve Bayes (NB) classifier [19], and
decision trees (DT) [20]. Convolutional neural networks (CNN) have gained popularity in
recent years due to their ability to perform automatic feature extraction [21–24], as well as
their broad use in research [25–28]. They are used to detect cancerous cells more rapidly
and effectively.
The mortality rates are rising to alarming levels, yet if patients are detected and treated
promptly, their chances of surviving are better than 95% [29–34]. Thus, this motivates us to
develop a model for the early diagnosis of skin cancer to save human lives. In this paper,
we present a novel multi-classification model, called the deep learning-based skin cancer
classification network (DSCC_Net), based on the CNN, that identifies the four types of
skin cancer, MEL, MN, BCC, and SCC, from dermoscopic images. Most of the research
studies [29–33] have indicated great performance in binary classification, i.e., differentiating
between benign and malignant skin cancer. However, no evidence has been found that
uses the DL models for the classification of the skin cancers MEL, BCC, MN, and SCC.
Additionally, DSCC_Net iwas also compared with six baseline classifiers: Vgg-19, Vgg-16,
ResNet-152, EfficientNet-B0, Inception-V3, and MobileNet. The major contributions of this
study are presented below:
• The novel proposed DSCC_Net model is designed to identify four different types of
skin cancer. The proposed model has the capability of extracting dominant features
from dermoscopy images that can assist in the accurate identification of the disease.
• In this study, we reduce the complexity of the model by decreasing the number of
trainable parameters to obtain a significant classifier.
• The CNN model’s accuracy is compromised as a result of the problem of class imbal-
ance in medical datasets. We overcome this issue by using an up-sampling technique,
SMOTE Tomek, to obtain concoction samples of the image at each class to gain en-
hanced accuracy.
• The Grad-CAM heat-map technique is utilized to illustrate the visible features of skin
cancer disease classification approaches.
• The proposed model achieved superior results, as compared to six baseline classifiers,
Vgg-19, ResNet-152, Vgg-16, MobileNet, Inception-V3, and EfficientNet-B0, in terms of
many evaluation metrics, i.e., accuracy, area under the curve (AUC), precision, recall,
loss, and F1 score.
• Additionally, the proposed model also produced significant results as compared to the
recent state-of-the-art classifiers.
This study is divided into the following section: Section 2 presents the literature review.
Materials and methods are discussed in Section 3. The experimental results and discussion
are presented in Section 4. This study is concluded in Section 5.

2. Literature Review
Extensive research has been conducted on the diagnosis of skin cancer to better assist
medical professionals in the process of detecting the disease at an earlier stage. Recent
research, on the other hand, has been focused on developing different artificial intelligence
algorithms to automate the diagnosis of several types of skin cancer. Table 1 presents the
summary of recent literature on skin cancer diagnosis using DL models.
Cancers 2023, 15, 2179 4 of 28

Table 1. Summary of the existing research studies for the diagnosis of skin cancer, using different
machine learning and DL models.

Ref Model Type Limitations Dataset Accuracy

The classification accuracy of the model may be
Two hybrid
[35] Benign vs. Melanoma. enhanced by using more advanced sampling ISBI 2016 88.02%
CNN Models
techniques and data preparation.
Spiking Melanoma vs.
[36] The model’s interpretability has to be improved. ISIC 2019 89.57%
Vgg-13 non-Melanoma.
CNN,
AlexNet, MEL, BCC, AKIEC, NV, There is a limited selection of lightweight
[37] HAM10000 92.25%
Vgg-16, BKL, DF, and VASC. networks and hyperparameters for evaluation.
Vgg-19
ISIC
The model’s segmentation performance is fragile 2016,
[29] Deep CNN Malignant vs. Benign. to occlusions in skin pictures, and it struggles ISIC 90.42%
with low-contrast skin disease images. 2017,
ISIC 2020
CNN and MEL, BCC, AKIEC, NV, Different models and datasets call for various
[38] HAM10000 86%
ResNet-50 BKL. hyperparameter settings.
To further enhance the model’s generality, a more
[39] DenseNet-201 MEL & non-MEL. ISIC 2019 76.08%
clinical dataset of skin-cancer cases is required.
Overall accuracy drops when there is a large gap
[30] MobileNet-V2 Malignant & benign. ISIC 2020 98.2%
between the data domain and the target domain.
The proposed model was trained and tested on
EfficientNets MEL, BCC, AKIEC, NV,
[40] an imbalanced dataset of skin cancer, and it HAM10000 87.91%
B0-B7 BKL, DF, and VASC.
affects the model performance.
Due to the small sample size of the datasets used
[31] DCNN Benign & malignant. in this study, local optimizations may HAM10000 91.93%
have been achieved.
ResNet-152,
The computational cost was significant, and the
SE-ResNeXt- MEL, BCC, AKIEC, NV,
[41] system did not take into account all ISIC 2018 93%
101, BKL, DF, and VASC.
possible skin cancers.
DenseNet-161
Classification persists, however, because the
MEL, BCC, AKIEC, NV,
[42] CNN model relies on a small quantity of training data HAM10000 78%
BKL, DF, and VASC.
and the hazy borders of skin disease pictures.
Due to the lack of adversarial training on other
[43] DenseNet-121 MEL, BCC, and AKIEC. skin cancer datasets, the method’s HAM10000 85%
model remains vulnerable.

Keerthana et al. [35] classified dermoscopy images as either benign or malignant

cancers using two new hybrid CNN models, including an SVM algorithm at the output
layer. The parameters extracted by the initial CNN model and the second CNN model are
combined and passed to the SVM classifier. The accuracy of the first hybrid model with
DenseNet-201 and MobileNet was 88.02%, whereas the accuracy of the second hybrid model
with DenseNet-201 and ResNet-50 was 87.43%. Deep spiking neural networks were applied
by Qasim Gilani et al. [36] to a total of 3670 melanoma images and 3323 non-melanoma
images taken from the ISIC 2019 dataset. Using the suggested spiking Vgg-13 model, they
attained an 89.57% accuracy and 90.07% F1-score, which was greater than that acquired
with Vgg-13 and AlexNet, with fewer trainable parameters. Using the HAM10000 dataset,
Kousis et al. [37] established 11 CNN architectures for several skin lesion classifications.
They also built a mobile android application, in which DenseNet-169 architecture was
applied that was relatively light, which identified the skin lesion as benign or malignant.
Finally, DenseNet-169 was the model that achieved the highest accuracy (92.25%) when
compared to other models, e.g., ResNet-50, Vgg-16, Inception-V3, etc. The second-highest
accuracy has been achieved by the DenseNet-121 model. In terms of mobile applications,
DenseNet-169 attained 91.10% accuracy. To accurately differentiate between malignant and
Cancers 2023, 15, 2179 5 of 28

benign melanoma, Kaur et al. [29] suggested an automatic melanoma classifier that was
based on a deep CNN. The main goal was to suggest a lightweight and less-complicated
deep CNN than other techniques, in order to efficiently identify melanoma skin tumors.
The ISIC datasets were used to obtain dermoscopic pictures for this study that included
several cancer samples such as ISIC 2016, ISIC 2017 and ISIC 2020. In terms of the ISIC
2016, 2017 and 2020 datasets, the suggested deep CNN classifier acquired accuracy rates of
81.41 %, 88.23 %, and 90.42 %.
Alwakid et al. [38] employed the CNN model and modified ResNet-50, which was
applied to a HAM10000 dataset. This analysis used an uneven sample of skin cancer.
Initially, the image’s quality was improved using ESRGAN, then the next step taken to
tackle the problem of class imbalance was the use of augmenting data. They achieved
the result by using the CNN and ResNet-50 models, which were 86% and 85.3% accurate,
respectively. Aljohani et al. [39] used CNN to perform binary classification for the detection
of melanoma skin tumors. They used the ISIC 2019 dataset to test various CNN architectures
for this purpose. The results of the experiment showed that GoogleNet achieved the
maximum level of performance on both the training and testing data, in which they
obtained 74.91% and 76.08% accuracies. Rashid et al. [30] used MobileNet-V2 to present a
deep transfer learning network for the classification of melanoma. The MobileNet-V2 was a
deep CNN that distinguished between malignant and benign skin lesions. The performance
of the suggested DL model had been analyzed using the dataset of ISIC 2020. To solve the
class imbalance problem, different data augmentation strategies were used. Ali et al. [40]
applied EfficientNets B0-B7 models to the HAM10000 dataset of dermatoscopic images.
The dataset contained 10015 images associated with seven different types of skin cancer,
such as actinic keratosis (AKIEC), dermatofibrosarcoma (DF), non-vascular (NV), BCC,
MEL, benign keratosis (BKL) and vascular skin lesions (VASC). Among the eight models,
the EfficientNet-B4 represented the greatest Top-1 and Top-2 accuracies. In this experiment,
the EfficientNet-B4 model achieved an 87% F1 score and 87.91% Top-1 accuracy.
Shahin-Ali et al. [31] used a deep CNN model by using the HAM10000 dataset. This
data contained 6705 benign images, 1113 malignant images, and 2197 unknown images
of lesions. The proposed model attained the highest training and testing accuracies, with
93.16 % and 91.93%, respectively. Furthermore, they balanced the dataset for both classes,
which increased the accuracy of categorization. On the same dataset, they also trained
several transfer learning models, but the results were not better than their proposed model.
Le et al. [44] introduced a transfer learning model that comprised ResNet-50 without the
use of a preprocessing stage or manual selection of features. All layers of the pre-trained
ResNet-50 were used for the training in Google Colab. Global average pooling and dropout
layers were employed to reduce overfitting. The images of the dataset were divided into
seven different categories and the proposed model attained 93% accuracy. Bajwa et al. [41]
created an ensemble model through the use of ResNet-152, SE-ResNeXt-101, DenseNet-161,
and NASNet, to classify seven types of skin cancer with 93% accuracy. The ensemble
was a technique of ML that merges the results of various distinctive learners to improve
classification performance. Nugroho et al. [42] used the HAM10000 dataset to create a
custom CNN for skin cancer identification. They used a scaled image with a resolution of
90 × 120 pixels. They achieved an 80% accuracy for training and 78% accuracy for testing.
Bassi et al. [45] used a DL technique that included transfer learning and fine-tuning.
They resized the dataset images with the resolution of 224 × 224 and used a fine-tuned
Vgg-16 model. They attained an accuracy of 82.8 %. Moldovan et al. [43] used a technique
that was based on DL and transfer learning, in which they applied the HAM10000 dataset.
The classification model was created in Python, utilizing the PyTorch library and a two-step
process for classifying images of skin cancer. The first step’s prediction model was 85.0%
accurate, and the second step’s prediction model was 75.0% accurate. Using dermoscopic
images, Çevik et al. [46] employed the VGGNET model that contained a powerful CNN
model to identify seven various kinds of disease. Images that were 600 × 450 pixels in size
were analyzed and resized to 400 × 300 pixels. Sklearn, Tensorflow and Keras machine
Cancers 2023, 15, 2179 6 of 28

learning packages all were used in this Python-coded application. They obtained a score
of 85.62 percent accuracy. Hasan et al. [47] developed the CNN-based detecting system
that used feature extraction techniques to extract features from dermoscopic pictures.
During the testing phase, they obtained an accuracy of detection of 89.5 %. However, the
detection accuracy was insufficient and needed to be improved. Furthermore, there was
overfitting between the testing and training stages, which was a flaw in that study. Saba
et al. [31] suggested a deep CNN that used three phases to detect skin lesions: first, the
color modification was used to improve contrast; second, a CNN approach was applied to
extract the borders of the lesion; third, transfer learning was applied to remove the deep
features. While the strategy produced good results for some datasets, the outcomes varied
depending on the dataset.
Using the dataset of ISIC 2018, Majtner et al. [48] created an ensemble of GoogleNet
and Vgg-16 models. The authors performed the data augmentation and normalized its color
to build the ensemble approaches they offered. The accuracy of the suggested method was
80.1%. Alquran et al. [33] introduced an image-processing-based approach for detecting,
extracting, and classifying tumors from dermoscopy pictures, which aided in the diagnosis
of benign and melanoma skin cancer significantly. The SVM classifier’s results showed an
accuracy of 92.1%. Lopez et al. [49] described a deep-learning-based strategy to handle
the problem of identifying a dermoscopic image that included a skin tumor as malignant
and benign, with a focus on the difficulty of skin cancer classification, especially initial
melanoma detection. The proposed solution employed the transfer learning approach that
was based on the VGGNet CNN architecture. The proposed method obtained an accuracy
level of 81.3% in the ISIC dataset, according to encouraging testing results. A linear classifier
was built by Kawahara et al. [50] using a dataset of 1300 pictures and features collected
by CNN to detect skin cancer. The method does not need skin lesion segmentation or
preprocessing. They conducted classifications of five and ten classes, and their respective
accuracy rates were 85.8% and 81.9%. Codella et al. [51] employed sparse coding, SVM,
and deep learning to obtain an accuracy of 93.1% when evaluating recorded photos from
the ISIC. These images were represented by bkl, mel, and nv. Krishnaraj et al. [52] designed
machine learning [53–56] classifiers that identified binary classes of cervical cancer, such
as adenosquamous carcinoma and SCC. They collected the dataset at the University of
California, Irvine (UCI) repository, and the Borderline-SMOTE approach was employed
to balance the unbalanced data. They obtained 98% accuracy through this dataset. Imran
et al. [57] proposed a model that was based on deep CNN by using different layers and filter
sizes. They used three different publicly available datasets: ISIC-2017, ISIC-2018, and ISIC-
2019. In the ISIC-2017 dataset, they employed 2750 images that consisted of three labels:
MEL, BKL, and NV. The ISIC-2018 dataset contains seven labels, in which 10,015 images
were used, whereas the ISIC-2019 dataset implemented eight labels that contain a total
number of 25,331 images. The accuracy rate of the ISIC-2017 dataset was 93.47%, while
88.75% and 89.58% accuracies were achieved by ISIC-2018 and ISIC-2019, respectively.
According to the above literature, it is extremely clear that a need still exists for a
model with the ability detect the four different types of skin cancer with greater accuracy
than current modalities. Although [29–31,39,47,49] performed a binary class classification
of skin cancer, many other researchers were not able to handle multiclass classification
with more successful outcomes. For multiclass skin cancer detection, the previous methods
proposed in [40–48] were also unsuccessful at attaining a greater accuracy. Automated skin
cancer classification in dermoscopic images is a challenging task due to high intraclass
variance and interclass visual similarity. Furthermore, the presence of external and inherent
artifacts and contrast between the affected and normal skin make it extremely difficult
for the multiclassification of skin cancers. The proposed method overcomes the existing
challenges, and effectively classifies the lesion into the four primary classes of skin cancer,
MEL, SCC, BCC, and MN, with high efficiency.
Cancers 2023, 15, 2179 7 of 28

3. Materials and Methods

This section presents the experimental procedure used to analyze the performance of
the proposed model, as well as six well-known deep CNN models, which include Vgg-19,
ResNet-152, Vgg-16, MobileNet, Inception-V3, and EfficientNet-B0.

3.1. Proposed Study Flow for the Diagnosis of Skin Cancer

When skin cells are exposed to UV radiation, their DNA is altered, which disrupts
the skin cell’s normal growth and results in skin cancer. To find skin cancer, researchers
frequently use dermoscopic images. DL algorithms are applied to enhance the accuracy of
the detection of skin cancers, such as MEL, BCC, MN, and SCC. Furthermore, if skin cancer
is diagnosed in its initial phase, health professionals have a better opportunity to prevent the
disease’s growth and start treatment on time. The medical field has changed significantly
as a direct result of the application of artificial intelligence and image processing. At
this time, image processing is employed for analysis in almost every area of the medical
field [58–60]. The community of researchers plays a significant role in the development
of intelligent automated systems for accurate and speedy evaluations, and contributes to
daily improvements of these systems [61–63].
For this study, we developed an automated system for the identification of skin cancers,
called DSCC_Net. This system was trained and tested on images of four main categories
of skin cancer: SCC, BCC, MN, and MEL. The input image’s size is fixed to a resolution
of 150 × 150 pixels. In addition, the dataset was used according to the data normalization
technique, in order to stop the model from being overfit. We also applied a technique called
the synthetic minority oversampling technique (SMOTE) Tomek, in order to tackle the issue
of an unequal distribution of datasets and to balance the number of samples within each
class [64]. The skin cancer dataset is separated into three distinct categories that included
training, testing, and validation sets. Furthermore, Figure 1 shows the work flow of the
proposed DSCC_Net for skin cancer. In comparison to [65–67], the training parameter’s
size is smaller. The experimental procedure was carried out for a maximum of 30 epochs.
After completion of all the epochs, the proposed DSCC_Net achieved the expected level of
accuracy throughout training and validation. The performance of the suggested method
(DSCC_Net) was analyzed and was differentiated from that of six pre-trained classifiers:
accuracy, loss, precision, recall, AUC, and F1-score. The Grad-CAM heat-map approach has
been employed to illustrate the visible aspects of skin cancer that underline the qualities
that affect its categorization. These characteristics have been used to highlight the
Cancers 2023, 15, x FOR PEER REVIEW 8 of aspects
29
that lead to the diagnosis of skin cancer.

Figure 1. Workflow of the proposed DSCC_Net model.

Figure 1. Workflow of the proposed DSCC_Net model.
3.2. Dataset Description
On the internet, there are many freely accessible datasets of dermoscopy images. Be-
cause skin cancer is so common all around the world, this research focused on dermos-
copy and photographic images of the disease. Images of four classes of skin cancer are
Cancers 2023, 15, 2179 8 of 28

3.2. Dataset Description

On the internet, there are many freely accessible datasets of dermoscopy images. Be-
cause skin cancer is so common all around the world, this research focused on dermoscopy
and photographic images of the disease. Images of four classes of skin cancer are shown
in Figure 2. The proposed DSCC_Net was trained and tested on three datasets that were
derived from three different resources. The ISIC-2020 Archive [68] is the world’s largest
collection of dermoscopic images of skin lesions that are available to the general public.
The images contained in this dataset were derived from a variety of different sources,
because multiple institutes contributed patient data of various ages. There are 33,126 der-
moscopic images, 579 images of malignant skin lesions, and 32,542 images of benign skin
lesions. These pictures were taken from more than 2000 patients. We used 579 images of
the melanoma class, and histopathology verified the diagnoses for all these images. The
remaining images are all part of a benign class that was not considered for this research.
Secondly, the HAM10000 database [69] includes 10,015 images that were produced by the
International Skin Image Collaboration in 2018. Based on this information, this dataset
consists of seven different data classes that identify the skin lesions. This database was
developed by two different groups: Queensland University in Australia, and the ViDIR
Group at the University of Vienna in Austria. In this dataset, we used 510 basal cell class
images, 1107 melanoma class images, and 2007 melanocytic nevi class images. These
dermoscopic images were taken from different populations, and the rest of the images were
not considered in this study. Thirdly, dermis.net [70] is the most comprehensive online
Cancers 2023, 15, x FOR PEER REVIEW 9 of 29
dermatology information source. It offers detailed images, differential diagnoses, and
additional information on nearly all skin conditions.

Figure 2. Original image samples of skin cancer extracted from three datasets.
datasets.

3.3. Using SMOTE Tomek to Balance Dataset

To resolve the issue of an unequal distribution of classes throughout the dataset, we
applied the up-sampling method. In this method, we obtain fusion samples for each class
by using the up-sampling algorithm SMOTE Tomek [64], as shown in Figure 3. This
method is first applied to the class of observations belonging to minority classes. SMOTE
Cancers 2023, 15, 2179 9 of 28

3.3. Using SMOTE Tomek to Balance Dataset

To resolve the issue of an unequal distribution of classes throughout the dataset, we
applied the up-sampling method. In this method, we obtain fusion samples for each class
by using the up-sampling algorithm SMOTE Tomek [64], as shown in Figure 3. This method
is first applied to the class of observations belonging to minority classes. SMOTE is one
of the most common and well-known oversampling methods used by data scientists to
generate artificial minority points in the minority class examples. The aim was to combine
SMOTE and Tomek techniques to improve the efficiency of dealing with the unbalanced
class. Synthetic points are generated by SMOTE through the implementation of the10KNN
Cancers 2023, 15, x FOR PEER REVIEW of 29
algorithm. The distribution of samples before the implementation of up-sampling is shown
in Table 2.

Figure 3. SMOTE Tomek generates samples of images

images to
to solve
solve the
the class
class imbalance
imbalance issue.
issue.

3.4. Proposed
Table 2. ImageModel
samples of skin cancer are distributed before up-sampling.
This section contains a complete description of the proposed DSCC_Net model.
No. of Classes Class Name No. of Images

3.4.1. Structure0 of the Proposed DSCC_NetBCC 510

1 MEL 1686
The CNN2 structure is designed after MN the human brain’s biological 2007
anatomy, and is
especially beneficial
3 for applications of computer
SCC vision, such as object recognition,
97 image
segmentation, and face detection. According to the concept of translation or space invari-
ance, a CNN Model
3.4. Proposed can identify the same feature in multiple images regardless of where it occurs
in the images [71–73]. In this study, we developed a robust DSCC_Net based on the CNN
This section contains a complete description of the proposed DSCC_Net model.
model to accurately classify skin cancer diseases. The DSCC_Net model consists of 5 con-
volutional blocks,
3.4.1. Structure andProposed
of the also includes a Rectified Linear Unit (ReLU) activation function, 1
DSCC_Net
dropout layer, 2 dense layers, and a softmax classification layer, as illustrated in Figure 4.
The CNN structure is designed after the human brain’s biological anatomy, and
Table 3 provides an overview of the dataset after the up-sampling technique, while a de-
is especially beneficial for applications of computer vision, such as object recognition,
tailed explanation of the suggested DSCC_Net model for the categorization of skin cancer
image segmentation, and face detection. According to the concept of translation or space
with the succeeding layers is presented in Table 4.

Table 3. Image samples of the Skin Cancer dataset are distributed after up-sampling.

No. of Classes Class Name No. of Images

0 BCC 2035
Cancers 2023, 15, 2179 10 of 28

invariance, a CNN can identify the same feature in multiple images regardless of where
it occurs in the images [71–73]. In this study, we developed a robust DSCC_Net based on
the CNN model to accurately classify skin cancer diseases. The DSCC_Net model consists
of 5 convolutional blocks, and also includes a Rectified Linear Unit (ReLU) activation
function, 1 dropout layer, 2 dense layers, and a softmax classification layer, as illustrated
in Figure 4. Table 3 provides an overview of the dataset after the up-sampling technique,
while a detailed explanation of the suggested DSCC_Net model for the categorization
Cancers 2023, 15, x FOR PEER REVIEW 11 of 29 of
skin cancer with the succeeding layers is presented in Table 4.

Figure 4. Architecture of proposed DSCC_Net used to classify skin cancer diseases.

Figure 4. Architecture of proposed DSCC_Net used to classify skin cancer diseases.
Table 4. The total number of parameters utilized in the proposed DSCC_Net model.
Table 3. Image samples of the Skin Cancer dataset are distributed after up-sampling.
Layer Type Output Shape Parameters
Input Layer
No. of Classes (None, 150,
Class Name 150, 3) No.0 of Images
Block01 (None, 150, 150, 8) 224
0
Block02
BCC
(None, 75, 75, 16) 1168
2035
1 MEL 1952
Block03 (None, 37, 37, 32) 4640
2 MN 2007
Block04 (None, 18, 18, 64) 18,496
3 SCC 2018
Block05 (None, 9, 9, 128) 73,856
Dropout_1 (None, 4, 4, 128) 0
Flatten (None, 2048) 0
Table 4. The total number of parameters utilized in the proposed DSCC_Net model.
Dense_1 (None, 512) 1,049,088
Layer Type
ReLu (None,Output
512) Shape 0 Parameters
Input Layer
Dense_2 (None,
(None, 4) 150, 150, 3) 2052 0
Output: SoftMax
Block01 (None, 4) 150, 150, 8)
(None, 0 224
Block02 Total Parameters: (None, 75, 75, 16) 1,149,524 1168
Block03 Trainable Parameters: (None, 37, 37, 32) 1,149,524 4640
Non-Trainable Parameters: 0
Block04 (None, 18, 18, 64) 18,496
Block05 (None, 9, 9, 128) 73,856
3.4.2. Convolutional Blocks of CNN Model
Dropout_1 (None, 4, 4, 128) 0
The convolutional block is the fundamental building component of the presented
Flatten
work, and each (None, 2048) 2D, a ReLU, and a pooling
convolutional block contains a convolutional 0
2D with a Dense_1
max value. The initializer for the kernel (None, 512)
layer LecunUniformV2 is created 1,049,088
to
assign layer ReLu
kernel weights. The gradient-vanishing issue512)
(None, is solved by using the activation
0
function ofDense_2
ReLU, which also simplifies the process for the
(None, 4) network to understand 2052 and
carry out its tasks
Output: in a timely way.
SoftMax (None, 4) 0
Total Parameters:
RGB channels are contained 1,149,524
in the input image. Our model’s initial layer is known
as the convolutional layer. This layer
Trainable initiates the process by applying filters, also known
Parameters: 1,149,524
Non-Trainable
as the kernel. The kernel’s Parameters:
size is dependent 0
on two values, as illustrated in Equation (1).
Cancers 2023, 15, 2179 11 of 28

3.4.2. Convolutional Blocks of CNN Model

The convolutional block is the fundamental building component of the presented
work, and each convolutional block contains a convolutional 2D, a ReLU, and a pooling
2D with a max value. The initializer for the kernel layer LecunUniformV2 is created to
assign layer kernel weights. The gradient-vanishing issue is solved by using the activation
function of ReLU, which also simplifies the process for the network to understand and
carry out its tasks in a timely way.
Cancers 2023, 15, x FOR PEER REVIEW RGB channels are contained in the input image. Our model’s initial layer is known 12 of as
29
the convolutional layer. This layer initiates the process by applying filters, also known as
the kernel. The kernel’s size is dependent on two values, as illustrated in Equation (1).

𝐹𝑖𝑙𝑡𝑒𝑟
Filter 𝑆𝑖𝑧𝑒(𝐹𝑆)
Size ( FS) = =
f w 𝑓×
𝑤×f h 𝑓ℎ (1)
(1)
where fw denotes the width of the filter and fh denotes the height of the filter. In our study,
where fw denotes the width of the filter and fh denotes the height of the filter. In our study,
weset
we setthe
thesize
sizeof
ofthe
thefilter
filtertoto3,3,so
soEquation
Equation(1)
(1)becomes
becomesFS
FS==33×× 3.
3. Feature
Feature identifiers
identifiers are
are
another name for these filters, and enable us to understand low-level visual aspects,
another name for these filters, and enable us to understand low-level visual aspects, such such
as edges
as edgesand
andcurves
curves[74].
[74].

3.4.3. Flattened
3.4.3. Flattened Layer
Layer
This layer
This layer is
is located
located among
among the
the convolution
convolution and
and dense
dense layers.
layers. Tensor
Tensor data
data types
types are
are
usedas
used asinputs
inputsfor forthe
the convolution
convolution layers,
layers, whereas
whereas dense
dense layers
layers demand
demand a one-dimen-
a one-dimensional
sional layout.
layout. So, the So, the flattened
flattened layer
layer was was applied
applied to translate
to translate the two-dimensional
the two-dimensional image
image repre-
representation into a one-dimensional input, which is presented
sentation into a one-dimensional input, which is presented in Figure 5. in Figure 5.

Figure5.5. The
Figure The fundamental
fundamental structure
structureof
ofthe
theflattened
flattenedlayer.
layer.

3.4.4.
3.4.4. Dropout
Dropout Layer
Layer
Our
Our modelutilized
model utilizedthis layer
this with
layer a dropout
with valuevalue
a dropout of 0.2.ofThis
0.2.value
This was
valueimplemented
was imple-
in order to prevent the overfitting of our proposed DSCC_Net
mented in order to prevent the overfitting of our proposed DSCC_Net model model [74]. The [74].
purpose
The
of this layer was to switch units on and off to decrease the model’s training
purpose of this layer was to switch units on and off to decrease the model’s training time andtime
the
complexity of the model. Consequently, the model learns the relevant features.
and the complexity of the model. Consequently, the model learns the relevant features.
3.4.5. Dense Block of Proposed DSCC_Net
3.4.5. Dense Block of Proposed DSCC_Net
In this research, we apply 2 dense blocks that consist of an activation function, which
In this research, we apply 2 dense blocks that consist of an activation function, which
is explained in the following sections.
is explained in the following sections.
ReLU Function
ReLU Function
Activation
Activation functions,
functions, which
which are
are mathematical
mathematical processes,
processes, determine
determine whether
whether or
or not
not
neural output should be passed on to the next layer. In general, they enable and disable
neural output should be passed on to the next layer. In general, they enable and disable the
network nodes.
the network Many
nodes. activation
Many functions
activation are used
functions are in DLin
used classifiers, but webut
DL classifiers, applied ReLU
we applied
ReLU due to its uncomplicated and time-saving computation. The activation of ReLU
works by replacing all negative outcomes with zero. This activation function was used on
the outputs of the convolutional layer.
Dense Layer
Cancers 2023, 15, 2179 12 of 28

due to its uncomplicated and time-saving computation. The activation of ReLU works
by replacing all negative outcomes with zero. This activation function was used on the
outputs of the convolutional layer.
Dense Layer
The dense layer accepts a single matrix as input and generates output according to
its characteristics. In these layers, images are identified and given a class label. A dense
layer with 4 neurons and a SoftMax activation function is responsible for generating the
model’s final output, which classifies the image into one of the four skin cancer disease
classes: MEL, BCC, SCC, and MN. SoftMax is applied after a few layers; this is a probability-
based activation function in which the total amount of classes represents the number of
neurons [69]. The total number of parameters is 1,149,524, which is split into two groups:
1,149,524 trainable parameters, and zero non-trainable parameters.

3.5. Model Evaluations

A confusion matrix was employed to check the performance of the model. Before
training the model, the dataset was separated into training and test sets. The model was
then evaluated using the test set. We applied a variety of metrics to evaluate the model’s
performance. The following evaluation metrics (see Equations (2)–(5)) are widely employed
to measure the effectiveness of the suggested DSCC_Net for skin cancer detection:

TP + TN
Accuracy = (2)
TP + FN + FP + TN

TP
Precision = (3)
TP + FP
TP
Recall = (4)
TP + FN

Precision × Recall
F1 − score = 2 × (5)
Precision + Recall

4. Results and Discussion

We compare DSCC_Net to the most recently developed deep networks in the following
section. The comparisons between the suggested DSCC_Net and six baseline deep networks
are discussed in this section.

4.1. Experimental Setup

Keras was used to implement a total of seven models: six baseline models and the
DSCC Net model. In addition, the programming of the approaches that are not directly con-
nected to convolutional networks was achieved in Python. The experiment was achieved
by using a computer running the Windows 10 operating system, equipped with an 11 GB
NVIDIA GPU and 32 GB of RAM.

4.2. Accuracy Compared with Other Models

Using the same dataset and SMOTE Tomek, we compared our suggested and recent
deep neural networks i.e., Vgg-19, ResNet-152, EfficientNet-B0, Vgg-16, Inception-V3, and
MobileNet. We also compared the proposed DSCC_Net before applying the SMOTE Tomek.
The system with SMOTE Tomek provides remarkable results for the proposed model. The
obtained accuracies for the proposed DSCC_Net model with SMOTE Tomek, DSCC_Net
without SMOTE Tomek, Vgg-16, ResNet-152, Vgg-19, MobileNet, EfficientNet-B0, and
Inception-V3 were 94.17%, 83.20%, 91.12%, 89.32%, 91.68%, 92.51%, 89.46%, and 91.82%,
respectively, as illustrated in Table 5. The significant improvement attained by the proposed
DSCC_Net model, applying the SMOTE Tomek, is illustrated in Figure 6.
Cancers 2023, 15, 2179 13 of 28

Table 5. Performance of the DSCC_Net model compared with baseline algorithms.

Classifiers Accuracy Precision Recall F1-Score AUC

Vgg-16 91.12% 92.09% 90.43% 91.13% 99.02%
Vgg-19 91.68% 92.23% 90.57% 91.71% 98.14%
MobileNet 92.51% 92.95% 91.40% 92.17% 98.75%
ResNet-152 89.32% 90.73% 88.21% 89.27% 98.74%
EfficientNet-B0 89.46% 90.21% 88.21% 89.31% 98.43%
Inception-V3 91.82% 92.28% 91.12% 91.76% 99.06%
Proposed Model (With SMOTE Tomek) 94.17% 94.28% 93.76% 93.93% 99.43%
Cancers 2023, 15, x FOR PEER REVIEW Proposed Model (Without 14 of 29
83.20% 85.01% 80.62% 58.09% 96.65%
SMOTE Tomek)

Figure 6.6.Remarkable
Figure Remarkableaccuracy
accuracy improvement
improvement withwith or without
or without SMOTESMOTE
TomekTomek
in the in the proposed
proposed model
model compared to other baseline deep networks; (a) Vgg-16, (b) Vgg-19, (c) EfficientNet-B0, (d)
compared to other baseline deep networks; (a) Vgg-16, (b) Vgg-19, (c) EfficientNet-B0, (d) ResNet-152,
ResNet-152, (e) Inception-V3, (f) MobileNet, (g) Proposed Model with SMOTE Tomek,
(e) Inception-V3, (f) MobileNet, (g) Proposed Model with SMOTE Tomek, and (h) Proposed Modeland (h) Pro-
posed Model
without SMOTE without SMOTE Tomek.
Tomek.

Table 5. Performance of the DSCC_Net model compared with baseline algorithms.

Classifiers Accuracy Precision Recall F1-score AUC

Vgg-16 91.12% 92.09% 90.43% 91.13% 99.02%
Cancers 2023, 15, x FOR PEER REVIEW 15 of 29

Proposed Model (Without SMOTE

Cancers 2023, 15, 2179 83.20% 85.01% 80.62% 58.09% 14 96.65%
of 28
Tomek)

4.3. AUC Comparison with Other Models

4.3. AUC Comparison with Other Models
As discussed earlier in this research, our suggested model is a deep CNN-based
As discussed
DSCC_Net modelearlier in this
containing research,
multiple our that
units suggested model is aefficient
are particularly deep CNN-based
in recognizing
DSCC_Net
various skin cancer classifications. We compared DSCC_Net with five in
model containing multiple units that are particularly efficient recognizing
baseline deep net-
various skin cancer classifications. We compared DSCC_Net with five baseline deep
works to validate our proposed DSCC_Net model. Six baseline models, ResNet-152, Vgg-
networks to validate our proposed DSCC_Net model. Six baseline models, ResNet-152,
19, EfficientNet-B0, Vgg-16, MobileNet and Inception-V3, achieved the AUC values of
Vgg-19, EfficientNet-B0, Vgg-16, MobileNet and Inception-V3, achieved the AUC values of
98.74%, 98.91%, 98.43%, 99.02%, 98.75% and 99.06% respectively. Figure 7 depicts that the
98.74%, 98.91%, 98.43%, 99.02%, 98.75% and 99.06% respectively. Figure 7 depicts that the
proposedDSCC_Net
proposed DSCC_Net withwith
SMOTESMOTE
TomekTomek and DSCC_Net
and DSCC_Net without
without SMOTE SMOTE
Tomek Tomek
achieved
achieved respective
respective 99.43% and99.43%
96.65%andAUC96.65% AUC
values whenvalues
usingwhen using theOn
the datasets. datasets. Onofthe
the basis thebasis
previous analysis, we conclud that the suggested model’s AUC results remain superior to su-
of the previous analysis, we conclud that the suggested model’s AUC results remain
perioroftoother
those those of other models.
models.

Figure7.7. Results
Figure Results of
of the
the proposed
proposed DSCC_Net
DSCC_Netmodel
modelwith
withand
andwithout
withoutup-sampling;
up-sampling;(a)(a)Vgg-16,
Vgg-16, (b)
(b) Vgg-19,
Vgg-19, (c)(c) EfficientNet-B0,(d)
EfficientNet-B0, (d) ResNet-152,
ResNet-152, (e)
(e)Inception-V3,
Inception-V3,(f)(f)
MobileNet, (g) Proposed
MobileNet, Model
(g) Proposed Model
withSMOTE
with SMOTETomek,Tomek, and
and (h)(h) Proposed
Proposed Model
Model without
without SMOTE
SMOTE Tomek.Tomek.
Cancers 2023,
Cancers 2023, 15,
15,2179
x FOR PEER REVIEW 16
15 of 29
of 28

4.4. Compared with Other Models Using

4.4. Using Precision
Precision
We examined
We examinedour oursuggested
suggested and
and existing
existing networks,
networks, suchsuch as ResNet-152,
as ResNet-152, Vgg-19,
Vgg-19, Vgg-
Vgg-16,
16, MobileNet,
MobileNet, EfficientNet-B0,
EfficientNet-B0, and and Inception-V3,
Inception-V3, on the
on the same
same dataset
dataset and and balanced
balanced it
it using
using SMOTE
SMOTE Tomek.
Tomek. TheThe system
system withwith
SMOTESMOTETomek Tomek generated
generated remarkable
remarkable results results
for the
for the proposed
proposed DSCC_Net.DSCC_Net. The proposed
The proposed DSCC_Net DSCC_Net
with andwith and without
without SMOTEattained
SMOTE Tomek Tomek
precision values ofvalues
attained precision 94.28%ofand 85.01%,
94.28% andbut ResNet-152,
85.01%, Vgg-16, EfficientNet-B0,
but ResNet-152, Vgg-19,
Vgg-16, EfficientNet-B0,
Inception-V3, and MobileNet
Vgg-19, Inception-V3, achieved precision
and MobileNet achievedvalues of 90.73%,
precision 92.09%,
values 90.12%, 92.09%,
of 90.73%, 92.23%,
92.28%, and 92.95%, respectively. As a result of this analysis, we found that
90.12%, 92.23%, 92.28%, and 92.95%, respectively. As a result of this analysis, we foundthe suggested
DSCC_Net ’s precision
that the suggested performance
DSCC_Net with SMOTE
’s precision Tomekwith
performance is superior
SMOTE and moreisconsistent
Tomek superior
compared to recent models,
and more consistent comparedas illustrated in Figure
to recent models, as8.illustrated in Figure 8.

Figure 8. Precision
Figure 8. Precision results
resultsof
ofthe
theproposed
proposedmodel,
model,DSCC_Net,
DSCC_Net, and
andother baseline
other models;
baseline (a)(a)
models; Vgg-16,
Vgg-
16, Vgg-19,
(b) (b) Vgg-19, (c) EfficientNet-B0,
(c) EfficientNet-B0, (d) ResNet-152,
(d) ResNet-152, (e) Inception-V3,
(e) Inception-V3, (f) MobileNet,
(f) MobileNet, (g) Proposed
(g) Proposed Model
Model
with with SMOTE
SMOTE Tomek, Tomek,
and (h) and (h) Proposed
Proposed Model without
Model without SMOTESMOTE
Tomek. Tomek.
Cancers 2023,
Cancers 2023, 15,
15, 2179
x FOR PEER REVIEW 1716of
of 29
28

4.5. Compared of DSCC_Net against Other Models Using Recall

Themodel’s
The model’sability
abilityto
toidentify
identify positive
positive samples
samples was
was evaluated
evaluated based on the recall
metric. High
High recall
recall values
values indicate
indicate that
that more
more positive
positive samples
samples were
were identified.
identified. The
The pro-
pro-
posed DSCC_Net model was compared to other baseline deep networks using a recall
curve, as illustrated in Figure 9. 9. The
The proposed
proposed DSCC_Net
DSCC_Net with and without SMOTE
Tomek,
Tomek, ResNet-152,
ResNet-152,EfficientNet-B0,
EfficientNet-B0, Vgg-19, Inception-V3,
Vgg-19, Vgg-16,
Inception-V3, and and
Vgg-16, MobileNet attained
MobileNet at-
the recall
tained thevalues
recall of 93.76%,
values 80.62%,80.62%,
of 93.76%, 88.21%,88.21%,
88.21%, 88.21%,
90.57%, 90.57%,
91.12%, 91.12%,
90.43% and 91.40%,
90.43% and
respectively.
91.40%, As a result
respectively. Asofa the above
result explanation,
of the the proposed
above explanation, themethod
proposedshows remarkable
method shows
recall performance.
remarkable recall performance.

Figure 9. The recall analysis

analysis measures
measures the
theproportion
proportionofoftrue
truepositive
positiveresults
resultscorrectly
correctlyidentified
identifiedbybya
a predictive model out of all actual positives; (a) Vgg-16, (b) Vgg-19, (c) EfficientNet-B0, (d)
predictive model out of all actual positives; (a) Vgg-16, (b) Vgg-19, (c) EfficientNet-B0, (d) ResNet-
ResNet-152,
152, (e) Inception-V3, (f) MobileNet, (g) Proposed Model with SMOTE Tomek,
(e) Inception-V3, (f) MobileNet, (g) Proposed Model with SMOTE Tomek, and (h) Proposed and (h) Proposed
Model
Model without
without SMOTESMOTE
Tomek.Tomek.
Cancers2023,
Cancers 2023,15,
15,2179
x FOR PEER REVIEW 18 of 28
17 of 29

4.6. F1-Score
4.6. F1-Score Comparison
Comparison with
with Recent
Recent Deep
DeepModel
Model
Theproposed
The proposedDSCC_Net
DSCC_Net model
model withwith
SMOTESMOTE
TomekTomek and DSCC_Net
and DSCC_Net without
without SMOTE
SMOTEachieved
Tomek Tomek theachieved thevalues
F1-score F1-score values and
of 93.93% of 93.93%
58.09%,and 58.09%, respectively.
respectively. Addi-
Additionally, the
tionally,
six themodels,
baseline six baseline models, ResNet-152,
ResNet-152, EfficientNet-B0,
EfficientNet-B0, Vgg-19, Inception-V3,
Vgg-19, Inception-V3, Mo-
MobileNet and
bileNet attained
Vgg-16, and Vgg-16, attained
the F1-score the of
values F1-score
89.27%,values
89.31%,of91.71%,
89.27%, 89.31%,
91.76%, 91.71%,
92.17%, and 91.76%,
91.13%,
respectively,
92.17%, and as illustrated
91.13%, in Figure
respectively, as 10. The suggested
illustrated in FigureDSCC_Net model attained
10. The suggested the
DSCC_Net
highest F1-scorethe
model attained value with F1-score
highest SMOTE Tomek shown
value with in Figure
SMOTE Tomek10. shown in Figure 10.

Figure 10. The

Figure10. Thevalue
valueof
ofthe
theF1-score
F1-scorebetween
betweenthe
theproposed
proposed model
modeland sixsix
and baseline models;
baseline (a) (a)
models; Vgg-16,
Vgg-
16,Vgg-19,
(b) (b) Vgg-19, (c) EfficientNet-B0,
(c) EfficientNet-B0, (d) ResNet-152,
(d) ResNet-152, (e) Inception-V3,
(e) Inception-V3, (f) MobileNet,
(f) MobileNet, (g) Proposed
(g) Proposed Model
Model
with with SMOTE
SMOTE Tomek, Tomek, and (h) Proposed
and (h) Proposed Model without
Model without SMOTE SMOTE
Tomek. Tomek.
Cancers 2023,
Cancers 2023, 15,
15, 2179
x FOR PEER REVIEW 19
18 of 29
of 28

4.7. Comparison of Proposed Model with Other Models Using Loss Loss
Loss functions are responsible for calculating the numerical difference between the
predicted and
predicted and actual values. In this study, a categorical cross-entropy method was utilized
to calculate the loss. When the model was trained using up-sampled photos, however, the
remarkable. The proposed DSCC_Net model with and without SMOTE
results were more remarkable.
Tomek
Tomek attained
attainedthetheloss
lossvalues
valuesofof0.1677%
0.1677% and 0.4332%,
and whereas
0.4332%, whereasResNet-152, EfficientNet-
ResNet-152, Efficient-
B0, Vgg-19,
Net-B0, MobileNet,
Vgg-19, MobileNet, Vgg-16, and
Vgg-16, Inception-V3
and Inception-V3achieved
achievedthe
theloss
lossvalues
values of
of 0.2613%,
0.2896%, 0.2353%,
0.2353%, 0.2525%,
0.2525%,0.2279
0.2279and and0.2189,
0.2189, respectively.
respectively. Figure
Figure 11 11 shows
shows the the major
major en-
enhancement
hancement in inthethe loss
loss value
value of of
thethe suggested
suggested DSCC_Net
DSCC_Net model
model with
with SMOTE
SMOTE Tomek.
Tomek.

Lossvalue
Figure 11. Loss valueofofthe
theproposed
proposedDSCC_Net
DSCC_Net model
model and
and other
other baseline
baseline models;
models; (a) Vgg-16,
(a) Vgg-16, (b)
Vgg-19,
(b) (c) (c)
Vgg-19, EfficientNet-B0,
EfficientNet-B0,(d)(d)
ResNet-152,
ResNet-152,(e)(e)Inception-V3,
Inception-V3,(f)
(f)MobileNet,
MobileNet,(g)
(g) Proposed
Proposed Model
with SMOTE
with SMOTE Tomek,
Tomek, andand (h)
(h)Proposed
ProposedModel
Modelwithout
withoutSMOTE
SMOTETomek.
Tomek.
Cancers 2023,
Cancers 2023, 15,
15, 2179
x FOR PEER REVIEW 2019of
of 29
28

4.8.
4.8. ROC
ROC Compared
Compared with
with Recent
Recent Model
Model
ROC
ROC iwa performed to evaluate the
iwa performed to evaluate effectiveness of
the effectiveness the diagnostic
of the diagnostic tests
tests and,
and, most
most
specifically, the reliability of the binary or multi-classifier. A receiver operating
specifically, the reliability of the binary or multi-classifier. A receiver operating charac- character-
istic (ROC)
teristic curve’s
(ROC) AUC
curve’s AUCis used to evaluate
is used the effectiveness
to evaluate of a classifier;
the effectiveness a higher
of a classifier; AUC
a higher
indicates
AUC that the
indicates classifier
that is moreiseffective.
the classifier Using the
more effective. dataset,
Using thewe evaluated
dataset, the reliability
we evaluated the
of our proposed
reliability DSCC_NetDSCC_Net
of our proposed model in model
terms ofin the ROC
terms curve,
of the ROC both withboth
curve, andwith
without
and
SMOTE SMOTE
without Tomek. Tomek.
This curve
Thiswas
curveused
wastoused
compare the proposed
to compare DSCC_Net
the proposed model,
DSCC_Net with
model,
and without
with SMOTE
and without Tomek,
SMOTE to sixto
Tomek, baseline models
six baseline on theonsame
models dataset.
the same The suggested
dataset. The sug-
DSCC_Net with and
gested DSCC_Net without
with SMOTE
and without Tomek,
SMOTE Vgg-19,
Tomek, Inception-V3,
Vgg-19, and MobileNet.
Inception-V3, Res-
and MobileNet.
Net-152, Vgg-16,
ResNet-152, and EfficientNet-B0
Vgg-16, and EfficientNet-B0 attained ROC values
attained of 0.9861,
ROC values 0.9145, 0.9145,
of 0.9861, 0.9711, 0.9742,
0.9711,
0.9818,
0.9742, 0.9778,
0.9818, 0.9759
0.9778,and 0.9572,
0.9759 and respectively, as shown
0.9572, respectively, asin Figurein12.
shown In the12.
Figure ROC curve,
In the ROC a
curve, a significant
significant enhancement
enhancement of the suggested
of the suggested DSCC_Net DSCC_Net
model’s model’s performance,
performance, with SMOTEwith
SMOTE can
Tomek, Tomek, can beinvisible
be visible Figurein12.Figure 12.

Figure 12. ROC

ROC curve
curve comparing
comparing the
the performance
performance of baseline
baseline models with the proposed
proposed DSCC_Net
model; (a) Vgg-16, (b) Vgg-19, (c) EfficientNet-B0, (d) ResNet-152, (e) Inception-V3,
Inception-V3, (f) MobileNet,
MobileNet,
(g) Proposed
(g) Proposed Model
Model with
with SMOTE
SMOTE Tomek,
Tomek, and
and (h)
(h) Proposed
ProposedModel
Modelwithout
withoutSMOTE
SMOTETomek.
Tomek.
Cancers 2023, 15, x FOR PEER REVIEW 21 of 29
Cancers 2023, 15, 2179 20 of 28

4.9. AU(ROC) Extension for Multi-Class Comparison against Recent Models

4.9. AU(ROC)
Figure 13Extension for Multi-Class
shows a comparison Comparison
between against Recent
the proposed Modelsmodel and six base-
DSCC_Net
line deep
Figuremodels using
13 shows the ROC curve’s
a comparison between extension. After DSCC_Net
the proposed balancing the dataset
model and bysix using
base-
line deep models
the SMOTE Tomekusing the ROCthe
technique, curve’s extension.
suggested Afterimproved
technique balancing significantly
the dataset by asusing
com-
the SMOTE
pared to theTomek technique,
six models, the can
which suggested
be seentechnique
in Figureimproved
13. The significantly as compared
significant impact of the
to the six models,
suggested which
DSCC_Net can be
model seen
was in Figure
observed 13. Theofsignificant
in terms the AUC for impact
both of the suggested
classes with and
DSCC_Net
without SMOTE modelTomek.
was observed in terms
The impacted of theinclude
classes AUC for both
class classesclass
0 (BCC), with1and
(MEL),without
class
SMOTE Tomek. The impacted classes include class 0 (BCC), class 1 (MEL),
2 (MN), and class 3 (SCC). These enhancements in AUC provide evidence that the feature class 2 (MN),
and class 3used
selection (SCC). These
by the enhancements
DSCC_Net in AUC
is accurate, and provide evidence
the SMOTE thatapproach
Tomek the feature selection
is also very
used by the DSCC_Net is accurate, and the SMOTE Tomek approach is also very useful.
useful.

Figure 13.
Figure 13. AU(ROC)
AU(ROC) curve
curveevaluation
evaluationwith
withextension
extensionfor
forthe
theproposed model
proposed model and other
and models;
other (a)
models;
Vgg-16, (b) Vgg-19, (c) EfficientNet-B0, (d) ResNet-152, (e) Inception-V3, (f) MobileNet, (g) Pro-
(a) Vgg-16, (b) Vgg-19, (c) EfficientNet-B0, (d) ResNet-152, (e) Inception-V3, (f) MobileNet, (g) Pro-
posedModel
posed Modelwith
withSMOTE
SMOTETomek,
Tomek,andand(h)
(h)Proposed
ProposedModel
Modelwithout
withoutSMOTE
SMOTETomek.
Tomek.
Cancers 2023, 15,
Cancers 2023, 15, 2179
x FOR PEER REVIEW 21 of
22 of 28
29

4.10. Comparison of
4.10. Comparison of DSCC_Net
DSCC_Net with
with Six
Six Models
Models Using
Using aa Confusion
Confusion Matrix
Matrix
To
To validate
validate our
our suggested
suggested DSCC_Net
DSCC_Net model
model with
with aa confusion
confusion matrix,
matrix, we
we compared
compared
it
it with
with six
six models. The use
models. The use of
of SMOTE Tomek results
SMOTE Tomek results in
in significant
significant improvements
improvements for
for the
the
DSCC_Net model, as presented in Figure
DSCC_Net model, as presented in Figure 14. 14.

Figure 14. Cont.

Cancers2023,
Cancers 2023,15,
15,2179
x FOR PEER REVIEW 23
22 of 29
of 28

Figure Usinga aconfusion

Figure 14. Using confusion matrix
matrix to compare
to compare the the proposed
proposed DSCC_Net
DSCC_Net with with
otherother deep net-
deep networks;
(a) Vgg-16,
works; (b) Vgg-19,
(a) Vgg-16, (c) EfficientNet-B0,
(b) Vgg-19, (d) ResNet-152,
(c) EfficientNet-B0, (e) Inception-V3,
(d) ResNet-152, (f) MobileNet,
(e) Inception-V3, (g) Pro-
(f) MobileNet,
posed
(g) ModelModel
Proposed with SMOTE Tomek,
with SMOTE and (h)
Tomek, Proposed
and ModelModel
(h) Proposed without SMOTE
without Tomek.
SMOTE Tomek.

The proposed method

The proposed methodaccurately
accuratelyclassifies
classifies 176
176 images
images outout of 190
of 190 totaltotal
imagesimages
in BCCin
BCC
cases,cases, whereas
whereas it misclassifies
it misclassifies 10 images
10 images as MN,as3 MN, 3 asand
as MEL, MEL, 1 asand 1 as
SCC. InSCC. In MN
MN classifi-
classification,
cation, 138 MN 138 MN images
images were correctly
were correctly identified
identified out of 164outtotal
of 164 total images,
images, while 13while
were
13 were misidentified as BCC, 9 as MEL images, and 4 as SCC images,
misidentified as BCC, 9 as MEL images, and 4 as SCC images, as illustrated in Figure 14. as illustrated in
Figure 14. The suggested method accurately identified 178 MEL images
The suggested method accurately identified 178 MEL images out of 179, whereas it mis- out of 179, whereas
it misclassified
classified one image
one image as BCC. as The
BCC. The DSCC_Net
DSCC_Net model model
correctlycorrectly
identifiedidentified
187 SCC 187 SCC
images
images outtotal
out of 188 of 188 totalwhile
images, images, while it misidentified
it misidentified one image as one
MN. image as MN.we
In addition, Inemployed
addition,
we
the employed
Grad-CAM theheatmap
Grad-CAM heatmap
approach to approach to visually
visually represent represent
the output the output
of our of our
suggested
suggested model. The objective of the heatmap is to show the relevant area
model. The objective of the heatmap is to show the relevant area of the skin that the model of the skin that
the model focuses on. Figure 15 illustrates the heatmap of
focuses on. Figure 15 illustrates the heatmap of the DSCC_Net model. the DSCC_Net model.
Cancers 2023, 15,
Cancers 2023, 15, 2179
x FOR PEER REVIEW 23of
24 of 28
29

Figure 15. Grad-CAM evaluation of the proposed DSCC_Net model for skin cancer diseases.
Figure 15. Grad-CAM evaluation of the proposed DSCC_Net model for skin cancer diseases.

4.11. Comparison of the Proposed Model with State-Of-The-Art

In this section, we compare our proposed DSCC_Net model with previous modern
studies [70–76]. Additionally, the proposed model is directly compared with the results
reported in
reported inthese
these[70–76]
[70–76]studies.
studies.Table
Table 6 presents
6 presents a comprehensive
a comprehensive analysis
analysis of theofproposed
the pro-
posed DSCC_Net
DSCC_Net model inmodel
termsinofterms
manyof many performance
performance evaluation evaluation metrics,
metrics, such such as recall,
as accuracy, accu-
F1-score, and
racy, recall, precision,
F1-score, andinprecision,
comparison with the recent
in comparison withstate-of-the-art studies.
the recent state-of-the-art studies.

Table 6.
Table 6. Comparison
Comparison of
of the
the DSCC_Net
DSCC_Net model
model with
with recent
recent state-of-the-art
state-of-the-artstudies.
studies.

Ref
Ref Year
Year Model
Model Datasets
Datasets Accuracy
Accuracy Recall
Recall Precision
Precision F1-Score
F1-Score
[70]
[70] 2023
2023 CNN
CNN ISIC-2017
ISIC-2017 92.00%
92.00% 91.90%
91.90% 91.65%
91.65% 91.99%
91.99%
[71] 2023 Vgg-13 ISIC-2019, Derm-IS 89.57% 90.70% 89.66% 89.65%
[71]
[72]
2023
2023
Vgg-13
Deep Belief Network
ISIC-2019, Derm-IS
HAM-10000
89.57%
93.00%
90.70%
92.91%
89.66%
92.45%
89.65%
92.65%
[72]
[73] 2023
2021 Deep Belief Network
ConvNet HAM-10000
ISIC-2018, Derm-IS 93.00% 86.90% 92.91%
86.14% 92.45%
87.47% 92.65%
-
[74] 2022 2D superpixels + RCNN HAM-10000 85.50% 83.40% 84.50% 85.30%
[73]
[75] 2021
2021 ConvNet
ResNeXt101 ISIC-2018,ISIC-2019
Derm-IS 86.90%
88.50% 86.14%
87.40% 87.47%
88.10% -
88.30%
[76]
[74] 2022
2022 SCDNet
2D superpixels + RCNN ISIC-2019
HAM-10000 92.91%
85.50% 92.18%
83.40% 92.19%
84.50% 92.18%
85.30%
ISIC-2020, Derm-IS,
Ours
[75] -
2021 DSCC_Net with SMOTE Tomek
ResNeXt101 ISIC-2019
HAM-10000 94.17% 87.40%
88.50% 94.28% 93.76%
88.10% 93.93%
88.30%
[76] 2022 SCDNet ISIC-2019 92.91% 92.18% 92.19% 92.18%
DSCC_Net 4.12.
with Discussions
SMOTE ISIC-2020, Derm-IS,
Ours - 94.17% 94.28% 93.76% 93.93%
Tomek The identificationHAM-10000
and categorization of a wide range of skin cancers may be accom-
plished with the use of dermoscopy photographs [32–35]. Our method offers a full view of
a4.12. Discussions
particular site, which enables us to identify the disease, as well as interior areas that have
beenTheinfected with it. Dermoscopy
identification is the most
and categorization reliable
of a wide [41]ofand
range time-effective
skin cancers may[52–59] ap-
be accom-
proach for determining if a lesion is a BCC, MEL, SCC, or MN. A computerized
plished with the use of dermoscopy photographs [32–35]. Our method offers a full view diagnostic
approach is required
of a particular to identify
site, which enables BCC,
us toMEL, SCC,the
identify anddisease,
MN, since the number
as well of confirmed
as interior areas that
cases of deadly
have been skinwith
infected cancer is continuallyisgrowing
it. Dermoscopy the most[62]. Dermoscopy
reliable images might [52–59]
[41] and time-effective be able
to automatically
approach differentiate
for determining if a between
lesion is athose
BCC,whoMEL, have
SCC, MEL and A
or MN. those who have other
computerized diag-
types
nostic of skin cancer,
approach by usingtomethods
is required identify from
BCC,the fieldSCC,
MEL, of DLand[64–72]. As a direct
MN, since result of
the number of
this, we developed a DSCC_Net model that is based on DL and
confirmed cases of deadly skin cancer is continually growing [62]. Dermoscopy imagesis capable of accurately
Cancers 2023, 15, 2179 24 of 28

diagnosing a wide variety of skin diseases. These diseases include BCC, MEL, SCC, and
MN, and the model enables dermatologists to begin treatment for their patients at an
earlier stage. The three publicly available benchmark datasets (i.e., ISIC 2020, HAM10000,
and DermIS) were used to evaluate the performance of the proposed DSCC_Net model.
The results of the proposed model were compared with six baseline models: ResNet-152,
Vgg-16, Vgg-19, Inception-V3, EfficientNet-B0, and MobileNet. The obtained image from
datasets is imbalanced as discussed in Table 2. The imbalanced class of the images affected
the performance of the model at the time of training [77–82]. To overcome these issues, we
used the SMOTE Tomek technique to increase the numbers of images in the minority class
of the datasets [49]. According to Figure 6, our proposed DSCC_Net model has received
sufficient training on the four subtypes of skin cancer (BCC, MEL, SCC, and MN), and it can
correctly identify occurrences of infection with these subtypes. Compared to the other six
baseline skin cancer classifiers, our DSCC_Net model performs much better in classifying
skin cancers, as discussed in Table 5. The DSCC_Net model using the SMOTE Tomek tech-
nique obtained an accuracy of 94.17%, regarding the categorization of dermoscopy pictures
of BCC, MEL, SCC, and MN. Additionally, the DSCC_Net model used without SMOTE
Tomek technique achieved an accuracy of 83.20%. On the other hand, the Vgg-16 model
attained an accuracy of 91.12%. Similarly, the Vgg-19 and MobileNet models achieved an
accuracy of 91.68% and 95.51%, respectively. The ResNet-152 model’s performance was
poor in skin cancer classification as compared to all baseline models. Furthermore, we
also provide the GRAD-CAM evaluation of the proposed DSCC_Net model for skin cancer
disease classification as shown in Figure 15.
Table 6 presents the classification performance of the proposed DSCC_Net model with
SOTA classifiers. Zhou et al. [70] proposed a DL model that achieved a classification accu-
racy of 0.92. Qasim et al. [71] designed a novel model, Vgg-13, for skin cancer identification.
They achieved an accuracy of skin cancer detection of 89.57%. A ConvNet net model that
focuses on the binary categorization of skin diseases was provided by Mijwil et al. [73].
This model was based on Inception-V3. By using this model, benign and malignant forms
of skin cancer are distinguished. The multiclassification of skin lesions was performed by
Afza et al. [74], by using 2D superpixels with ResNet-50, and they reached an accuracy of
85.50%. In addition, Khan et al. [75] attained a precision of 88.50% when performing the
multiclassification of skin cancer. When compared to other approaches that are considered
to be SOTA, the DSCC_Net model obtained an impressive accuracy of 94.71%.

5. Conclusions
In this study, the proposed DSCC_Net model, used for identifying the four forms
of skin cancer (BCC, MEL, SCC, and MN), was developed and evaluated. Today, these
skin cancer diseases are rapidly spreading and affect communities globally. Many deaths
have occurred because of improper and slow testing procedures, limited facilities, and the
lack of diagnosis of skin cancer at an early stage. Due to a large number of cases, a rapid
and effective testing procedure is necessary. We proposed a DSCC_Net model to identify
the four types of skin cancer diseases. Each convolutional block of the modified structure
was generated using multiple layers and was applied in order to classify early-stage skin
cancers. The SMOTE Tomek algorithm was used to generate samples that were used to
solve dataset imbalance problems and to maintain a balance in the number of samples for
each class. Grad-CAM displays a heat map of class activation to illustrate the operation of
the CNN layer. Our proposed DSCC_Net model achieved 94.17% accuracy, 93.76% recall,
93.93% F1-score, 94.28% precision, and 99.42% AUC. So, it is concluded that DSCC_Net
model can play a significant role as a supporting hand for the medical professional. The
limitation of the study is that our proposed DSCC_Net model is suitable for only fair-
skinned individuals. Individuals with dark skin were not considered in this study. The
reason is that the publicly available datasets used in this work contain skin cancer images
of fair-toned skin. In the future, we will combine blockchain and federated learning with a
Cancers 2023, 15, 2179 25 of 28

deep attention module to obtain more favorable results in classifying skin cancer, as well as
skin infections.

Author Contributions: Conceptualization, M.T., A.N. and H.M.; methodology, M.T., A.N. and H.M;
validation, R.A.N., J.T., S.-W.L. and H.M.; formal analysis, A.N. and S.-W.L.; investigation, A.N. and
R.A.N.; resources, A.N., J.T. and H.M; data curation, R.A.N.; writing—original draft preparation,
A.N., H.M; writing—review and editing, A.N., H.M. and R.A.N.; visualization, J.T., H.M; supervision,
H.M., R.A.N. and S.-W.L.; funding acquisition, S.-W.L. All authors have read and agreed to the
published version of the manuscript.
Funding: This work was supported by a national research foundation (NRF) grant funded by the
Ministry of Science and ICT (MSIT), South Korea through the Development Research Program
(NRF2021R1I1A2059735 and NRF2022R1G1A1010226).
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. World Health Organization Radiation: Ultraviolet (UV) Radiation and Skin Cancer|How Common Is Skin Cancer. Avail-
able online: https://www.who.int/news-room/q-a-detail/radiation-ultraviolet-(uv)-radiation-and-skin-cancer# (accessed on
2 March 2023).
2. Piccialli, F.; Di Somma, V.; Giampaolo, F.; Cuomo, S.; Fortino, G. A survey on deep learning in medicine: Why, how and when?
Inf. Fusion 2021, 66, 111–137. [CrossRef]
3. Navid, R.; Ashourian, M.; Karimifard, M.; Estrela, V.V.; Loschi, H.J.; Nascimento, D.D.; França, R.P.; Vishnevski, M. Computer-
aided diagnosis of skin cancer: A review. Curr. Med. Imaging 2020, 16, 781–793.
4. Ahmad, N.; Farooq, M.S.; Khelifi, A.; Abid, A. Malignant melanoma classification using deep learning: Datasets, performance
measurements, challenges and opportunities. IEEE Access 2020, 8, 110575–110597.
5. O’Sullivan, D.E.; Brenner, D.R.; Demers, P.A.; Villeneuve, P.J.; Friedenreich, C.M.; King, W.D. Indoor tanning and skin cancer in
Canada: A meta-analysis and attributable burden estimation. Cancer Epidemiol. 2019, 59, 1–7. [CrossRef] [PubMed]
6. Walters-Davies, R. Skin cancer: Types, diagnosis and prevention. Evaluation 2020, 14, 34.
7. Hodis, E. The Somatic Genetics of Human Melanoma. Ph.D. Thesis, Harvard University, Cambridge, MA, USA, 2018.
8. Nathan, N.; Hubbard, M.; Nordmann, T.; Sperduto, P.W.; Clark, H.B.; Hunt, M.A. Effect of gamma knife radiosurgery and
programmed cell death 1 receptor antagonists on metastatic melanoma. Cureus 2017, 9, e1943.
9. Ahmad, N.; Anees, T.; Naqvi, R.A.; Loh, W.-K. A comprehensive analysis of recent deep and federated-learning-based method-
ologies for brain tumor diagnosis. J. Pers. Med. 2022, 12, 275.
10. Rogers, H.W.; Weinstock, M.A.S.; Feldman, R.; Coldiron, B.M. Incidence estimate of non-melanoma skin cancer (keratinocyte
carcinomas) in the US population 2012. JAMA Dermatol. 2015, 151, 1081–1086. [CrossRef] [PubMed]
11. Bomm, L.; Benez, M.D.V.; Maceira, J.M.P.; Succi, I.C.B.; Scotelaro, M.D.F.G. Biopsy guided by dermoscopy in cutaneous pigmented
lesion-case report. An. Bras. Dermatol. 2013, 88, 125–127. [CrossRef] [PubMed]
12. Kato, J.; Horimoto, K.; Sato, S.; Minowa, T.; Uhara, H. Dermoscopy of melanoma and non-melanoma skin cancers. Front. Med.
2019, 6, 180. [CrossRef]
13. Haenssle, H.A.; Fink, C.; Schneiderbauer, R.; Toberer, F.; Buhl, T.; Blum, A.; Kalloo, A.; Hadj, H.A.B.; Thomas, L.; Enk, A.; et al.
Reader study level-I and level-II Groups, Man against machine: Diagnostic performance of a deep learning convolutional neural
network for dermoscopic melanoma recognition in comparison to 58 dermatologists. Ann. Oncol. 2018, 29, 1836–1842. [CrossRef]
14. Ibrahim, H.; El-Taieb, M.; Ahmed, A.; Hamada, R.; Nada, E. Dermoscopy versus skin biopsy in diagnosis of suspicious skin
lesions. Al-Azhar Assiut Med. J. 2017, 15, 203. [CrossRef]
15. Duggani, K.; Nath, M.K. A technical review report on deep learning approach for skin cancer detection and segmentation. Data
Anal. Manag. Proc. ICDAM 2021, 54, 87–99.
16. Carli, P.; Quercioli, E.; Sestini, S.; Stante, M.; Ricci, L.; Brunasso, G.; DeGiorgi, V. Patternanalysis, notsimplifiedalgorithms, isthe
most reliable method for teaching dermoscopy for melanoma diagnosis to residents in dermatology. Br. J. Dermatol. 2003, 148,
981–984. [CrossRef] [PubMed]
17. Carrera, C.; Marchetti, M.A.; Dusza, S.W.; Argenziano, G.; Braun, R.P.; Halpern, A.C.; Jaimes, N.; Kittler, H.J.; Malvehy, J.; Menzies,
S.W.; et al. Validity and reliability of dermoscopic criteria used to differentiate nevi from melanoma: A web-based international
dermoscopy society study. JAMA Dermatol. 2016, 152, 798–806. [CrossRef] [PubMed]
18. Celebi, M.E.; Kingravi, H.A.; Uddin, B.; Iyatomi, H.; Aslandogan, Y.A.; Stoecker, W.V.; Moss, R.H. A methodological approach to
the classification of dermoscopy images. Comput. Med. Imaging Graph. 2007, 31, 362–373. [CrossRef]
Cancers 2023, 15, 2179 26 of 28

19. Maglogiannis, I.; Doukas, C.N. Overview of advanced computer vision systems for skin lesions characterization. IEEE Trans. Inf.
Technol. Biomed. 2009, 13, 721–733. [CrossRef]
20. Celebi, M.E.; Iyatomi, H.; Stoecker, W.V.; Moss, R.H.; Rabinovitz, H.S.; Argenziano, G.; Soyer, H.P. Automatic detection of
blue-white veil and related structures in dermoscopy images. Comput. Med. Imaging Graph. 2008, 32, 670–677. [CrossRef]
[PubMed]
21. Hassaan, M.; Anees, T.; Din, M.; Ahmad, N. CDC_Net: Multi-classification convolutional neural network model for detection
of COVID-19, pneumothorax, pneumonia, lung Cancer, and tuberculosis using chest X-rays. Multimed. Tools Appl. 2022, 82,
13855–13880.
22. Lu, S.; Lu, Z.; Zhang, Y.D. Pathological brain detection based on AlexNet and transfer learning. J. Comput. Sci. 2019, 30, 41–47.
[CrossRef]
23. Ahmad, N.; Anees, T.; Ahmed, K.T.; Naqvi, R.A.; Ahmad, S.; Whangbo, T. Deep learned vectors’ formation using auto-correlation,
scaling, and derivations with CNN for complex and huge image retrieval. Complex Intell. Syst. 2022, 4, 1–23.
24. Sajjad, M.; Khan, S.; Muhammad, K.; Wu, W.; Ullah, A.; Baik, S.W. Multi-grade brain tumor classification using deep CNN with
extensive data augmentation. J. Comput. Sci. 2019, 30, 174–182. [CrossRef]
25. Alom, M.Z.; Aspiras, T.; Taha, T.M.; Asari, V.K. Skin cancer segmentation and classification with improved deep convolutional
neural network. In Medical Imaging 2020: Imaging Informatics for Healthcare, Research, and Applications; International Society for
Optics and Photonics: Bellingham, WA, USA, 2020; Volume 11318, p. 1131814.
26. Esteva, A.; Kuprel, B.; Novoa, R.A.; Ko, J.; Swetter, S.M.; Blau, H.M.; Thrun, S. Dermatologist-level classification of skin cancer
with deep neural networks. Nature 2017, 542, 115–118. [CrossRef]
27. Polat, K.; Koc, K.O. Detection of skin diseases from dermoscopy image using the combination of convolutional neural network
and one-versus-all. J. Artif. Intell. Syst. 2020, 2, 80–97. [CrossRef]
28. Ratul, M.A.R.; Mozaffari, M.H.; Lee, W.S.; Parimbelli, E. Skin lesions classification using deep learning based on dilated
convolution. BioRxiv 2020, 860700. [CrossRef]
29. Ranpreet, K.; GholamHosseini, H.; Sinha, R.; Lindén, M. Melanoma classification using a novel deep convolutional neural
network with dermoscopic images. Sensors 2022, 22, 1134.
30. Javed, R.; Ishfaq, M.; Ali, G.; Saeed, M.R.; Hussain, M.; Alkhalifah, T.; Alturise, F.; Samand, N. Skin cancer disease detection using
transfer learning technique. Appl. Sci. 2022, 12, 5714.
31. Shahin, A.; Miah, S.; Haque, J.; Rahman, M.; Islam, K. An enhanced technique of skin cancer classification using deep convolutional
neural network with transfer learning models. Mach. Learn. Appl. 2021, 5, 100036.
32. Tanzila, S.; Khan, M.A.; Rehman, A.; Marie-Sainte, S.L. Region extraction and classification of skin cancer: A heterogeneous
framework of deep CNN features fusion and reduction. J. Med. Syst. 2019, 43, 289.
33. Hiam, A.; Qasmieh, I.A.; Alqudah, A.M.; Alhammouri, S.; Alawneh, E.; Abughazaleh, A.; Hasayen, F. The melanoma skin
cancer detection and classification using support vector machine. In Proceedings of the 2017 IEEE Jordan Conference on Applied
Electrical Engineering and Computing Technologies (AEECT), Amman, Jordania, 11–13 October 2017; pp. 1–5.
34. Hardik, N.; Singh, S.P. Deep learning solutions for skin cancer detection and diagnosis. Mach. Learn. Health Care Perspect. Mach.
Learn. Healthc. 2020, 13, 159–182.
35. Duggani, K.; Venugopal, V.; Nath, M.K.; Mishra, M. Hybrid convolutional neural networks with SVM classifier for classification
of skin cancer. Biomed. Eng. Adv. 2023, 5, 100069.
36. Gilani, Q.; Syed, S.T.; Umair, M.; Marques, O. Skin Cancer Classification Using Deep Spiking Neural Network. J. Digit. Imaging
2023, 1–11.
37. Ioannis, K.; Perikos, I.; Hatzilygeroudis, I.; Virvou, M. Deep learning methods for accurate skin cancer recognition and mobile
application. Electronics 2022, 11, 1294.
38. Ghadah, A.; Gouda, W.; Humayun, M.; Sama, N.U. Melanoma Detection Using Deep Learning-Based Classifications. Healthcare
2022, 10, 2481.
39. Khalil, A.; Turki, T. Automatic Classification of Melanoma Skin Cancer with Deep Convolutional Neural Networks. AI 2022, 3,
512–525.
40. Karar, A.; Shaikh, Z.A.; Khan, A.A.; Laghari, A.A. Multiclass skin cancer classification using EfficientNets—A first step towards
preventing skin cancer. Neurosci. Inform. 2022, 2, 100034. [CrossRef]
41. Naseer, B.M.; Muta, K.; Malik, M.I.; Siddiqui, S.A.; Braun, S.A.; Homey, B.; Dengel, A.; Ahmed, S. Computer-aided diagnosis of
skin diseases using deep neural networks. Appl. Sci. 2020, 10, 2488.
42. Adi, N.A.; Slamet, I.S. Skins cancer identification system of HAMl0000 skin cancer dataset using convolutional neural network.
AIP Conf. Proc. 2019, 2202, 020039.
43. Moldovan, D. Transfer learning based method for two-step skin cancer images classification. In Proceedings of the 2019 E-Health
and Bioengineering Conference (EHB), Iasi, Romania, 21–23 November 2019; pp. 1–4.
44. Le Duyen, N.T.; Hieu, X.L.; Lua, T.N.; Hoan, T.N. Transfer learning with class-weighted and focal loss function for automatic skin
cancer classification. arXiv 2020, arXiv:2009.05977.
45. Saksham, B.; Gomekar, A. Deep learning diagnosis of pigmented skin lesions. In Proceedings of the 2019 10th International
Conference on Computing, Communication and Networking Technologies (ICCCNT), Kanpur, India, 6–8 July 2019; pp. 1–6.
Cancers 2023, 15, 2179 27 of 28

46. Emrah, Ç.; Zengin, K. Classification of skin lesions in dermatoscopic images with deep convolution network. Avrupa Bilim Ve
Teknol. Derg. 2019, 6, 309–318.
47. Hasan, M.; Barman, S.D.; Islam, S.; Reza, A.W. Skin cancer detection using convolutional neural network. In Proceedings of the
2019 5th International Conference on Computing and Artificial Intelligence, Bali, Indonesia, 19–22 April 2019; pp. 254–258.
48. Tomáš, M.; Bajić, B.; Yildirim, S.; Hardeberg, J.Y.; Lindblad, J.; Sladoje, N. Ensemble of convolutional neural networks for
dermoscopic images classification. arXiv 2018, arXiv:1808.05071.
49. Lopez; Romero, A.; Giro-i-Nieto, X.; Burdick, J.; Marques, O. Skin lesion classification from dermoscopic images using deep
learning techniques. In Proceedings of the 2017 13th IASTED International Conference on Biomedical Engineering (BioMed),
Innsbruck, Austria, 20–21 February 2017; pp. 49–54.
50. Jeremy, K.; BenTaieb, A.; Hamarneh, G. Deep features to classify skin lesions. In Proceedings of the 2016 IEEE 13th International
Symposium on Biomedical Imaging (ISBI), Prague, Czech Republic, 13–16 April 2016; pp. 1397–1400.
51. Noel, C.; Cai, J.; Abedini, M.; Garnavi, R.; Halpern, A.; Smith, J.R. Deep learning, sparse coding, and SVM for melanoma
recognition in dermoscopy images. In Proceedings of the Machine Learning in Medical Imaging: 6th International Workshop,
MLMI 2015, Held in Conjunction with MICCAI 2015, Munich, Germany, 5 October 2015; pp. 118–126.
52. Chadaga, K.; Prabhu, S.; Sampathila, N.; Chadaga, R.; Sengupta, S. Predicting cervical cancer biopsy results using demographic
and epidemiological parameters: A custom stacked ensemble machine learning approach. Cogent Eng. 2022, 9, 2143040. [CrossRef]
53. Sampathila, N.; Chadaga, K.; Goswami, N.; Chadaga, R.P.; Pandya, M.; Prabhu, S.; Bairy, M.G.; Katta, S.S.; Bhat, D.; Upadya, S.P.
Customized Deep Learning Classifier for Detection of Acute Lymphoblastic Leukemia Using Blood Smear Images. Healthcare
2022, 10, 1812. [CrossRef] [PubMed]
54. Krishnadas, P.; Chadaga, K.; Sampathila, N.; Rao, S.; Prabhu, S. Classification of Malaria Using Object Detection Models.
Informatics 2022, 9, 76. [CrossRef]
55. Acharya, V.; Dhiman, G.; Prakasha, K.; Bahadur, P.; Choraria, A.; Prabhu, S.; Chadaga, K.; Viriyasitavat, W.; Kautish, S. AI-assisted
tuberculosis detection and classification from chest X-rays using a deep learning normalization-free network model. Comput.
Intell. Neurosci. 2022, 2022, 2399428. [CrossRef]
56. Khanna, V.V.; Chadaga, K.; Sampathila, N.; Prabhu, S.; Bhandage, V.; Hegde, G.K. A Distinctive Explainable Machine Learning
Framework for Detection of Polycystic Ovary Syndrome. Appl. Syst. Innov. 2023, 6, 32. [CrossRef]
57. Imran, I.; Younus, M.; Walayat, K.; Kakar, M.U.; Ma, J. Automated multi-class classification of skin lesions through deep
convolutional neural network with dermoscopic images. Comput. Med. Imaging Graph. 2021, 88, 101843.
58. WHO. Gastrointestinal Cancer. 2020. Available online: https://www.who.int/news-room/fact-sheets/detail/cancer (accessed
on 2 March 2023).
59. Yogapriya, J.; Venkatesan Chandran, M.G.; Sumithra, P.; Anitha, P.; Jenopaul, C.; Dhas, S.G. Gastrointestinal tract disease
classification from wireless endoscopy images using pretrained deep learning model. Comput. Math. Methods Med. 2021, 2021,
5940433. [CrossRef] [PubMed]
60. Laith, A.; Fadhel, M.A.; Al-Shamma, O.; Zhang, J.; Santamaría, J.; Duan, Y.; Oleiwi, S.R. Towards a better understanding of
transfer learning for medical imaging: A case study. Appl. Sci. 2020, 10, 4523.
61. Yixuan, Y.; Li, B.; Meng, M.Q.-H. Bleeding frame and region detection in the wireless capsule endoscopy video. IEEE J. Biomed.
Health Inform. 2015, 20, 624–630.
62. Naveen, S.; Zverev, V.I.; Keller, H.; Pane, S.; Egolf, P.W.; Nelson, B.J.; Tishin, A.M. Magnetically guided capsule endoscopy. Med.
Phys. 2017, 44, e91–e111.
63. Benjamin, J.S.; Ferdinand, J.R.; Clatworthy, M.R. Using single-cell technologies to map the human immune system—Implications
for nephrology. Nat. Rev. Nephrol. 2020, 16, 112–128.
64. Hui, H.; Wang, W.-Y.; Mao, B.-H. Borderline-SMOTE: A new over-sampling method in imbalanced data sets learning. In
Proceedings of the Advances in Intelligent Computing: International Conference on Intelligent Computing, ICIC 2005, Hefei,
China, 23–26 August 2005; pp. 878–887.
65. Vasileios, C.; Tsiligiri, A.; Hadjileontiadis, L.J.; Liatsos, C.N.; Mavrogiannis, C.C.; Sergiadis, G.D. Ulcer detection in wireless
capsule endoscopy images using bidimensional nonlinear analysis. In Proceedings of the XII Mediterranean Conference on
Medical and Biological Engineering and Computing 2010, Chalkidiki, Greece, 27–30 May 2010; pp. 236–239.
66. Ayidzoe, A.; Mighty; Yu, Y.; Mensah, P.K.; Cai, J.; Adu, K.; Tang, Y. Gabor capsule network with preprocessing blocks for the
recognition of complex images. Mach. Vis. Appl. 2021, 32, 91. [CrossRef]
67. Mohapatra, S.; Nayak, J.; Mishra, M.; Pati, G.K.; Naik, B.; Swarnkar, T. Wavelet transform and deep convolutional neural
network-based smart healthcare system for gastrointestinal disease detection. Interdiscip. Sci. Comput. Life Sci. 2021, 13, 212–228.
[CrossRef] [PubMed]
68. The ISIC 2020 Challenge Dataset. Available online: https://challenge2020.isic-archive.com/ (accessed on 2 March 2023).
69. Philipp, T.; 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.
70. Dermtology Information System. Available online: http://www.dermis.net (accessed on 2 March 2023).
71. Sen, W.; Xing, Y.; Zhang, L.; Gao, H.; Zhang, H. Deep convolutional neural network for ulcer recognition in wireless capsule
endoscopy: Experimental feasibility and optimization. Comput. Math. Methods Med. 2019, 2019, 7546215.
Cancers 2023, 15, 2179 28 of 28

72. Nature. Olympus. The Endocapsule 10 System. Olympus Homepage. 2021. Available online: https://www.olympus-europa.
com/medical/en/Products--and--Solutions/Products/Product/ENDOCAPSULE-10-System.html (accessed on 2 March 2023).
73. Fushuan, W.; David, A.K. A genetic algorithm based method for bidding strategy coordination in energy and spinning reserve
markets. Artif. Intell. Eng. 2001, 15, 71–79.
74. Hassaan, M.; Farooq, M.S.; Khelifi, A.; Abid, A.; Qureshi, J.N.; Hussain, M. A comparison of transfer learning performance versus
health experts in disease diagnosis from medical imaging. IEEE Access 2020, 8, 139367–139386.
75. Ling, W.; Wang, X.; Fu, J.; Zhen, L. A Novel Probability Binary Particle Swarm Optimization Algorithm and its Application.
J. Softw. 2008, 3, 28–35.
76. Yufei, Z.; Koyuncu, C.; Lu, C.; Grobholz, R.; Katz, I.; Madabhushi, A.; Janowczyk, A. Multi-site cross-organ calibrated deep
learning (MuSClD): Automated diagnosis of non-melanoma skin cancer. Med. Image Anal. 2023, 84, 102702.
77. Alam, T.M.; Shaukat, K.; Khan, W.A.; Hameed, I.A.; Almuqren, L.A.; Raza, M.A.; Aslam, M.; Luo, S. An Efficient Deep
Learning-Based Skin Cancer Classifier for an Imbalanced Dataset. Diagnostics 2022, 12, 2115. [CrossRef] [PubMed]
78. Manash, E.B.K.; Suhasini, A.; Satyala, N. Intelligent skin cancer diagnosis using adaptive k-means segmentation and deep
learning models. Concurr. Comput. Pract. Exp. 2023, 35, e7546.
79. Mijwil, M.M. Skin cancer disease images classification using deep learning solutions. Multimed. Tools Appl. 2021, 80, 26255–26271.
[CrossRef]
80. Farhat, A.; Sharif, M.; Mittal, M.; Khan, M.A.; Hemanth, D.J. A hierarchical three-step superpixels and deep learning framework
for skin lesion classification. Methods 2022, 202, 88–102.
81. Khan, A.M.; Akram, T.; Zhang, Y.-D.; Sharif, M. Attributes based skin lesion detection and recognition: A mask RCNN and
transfer learning-based deep learning framework. Pattern Recognit. Lett. 2021, 143, 58–66. [CrossRef]
82. Naeem, A.; Anees, T.; Fiza, M.; Naqvi, R.A.; Lee, S.-W. SCDNet: A Deep Learning-Based Framework for the Multiclassification of
Skin Cancer Using Dermoscopy Images. Sensors 2022, 22, 5652. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.