THE USA JOURNALS
THE AMERICAN JOURNAL OF MEDICAL SCIENCES AND PHARMACEUTICAL RESEARCH
(ISSN
–
2689-1026)
VOLUME 06 ISSUE12
92
https://www.theamericanjournals.com/index.php/tajmspr
PUBLISHED DATE: - 21-12-2024
DOI: -
https://doi.org/10.37547/TAJMSPR/Volume06Issue12-10
PAGE NO.: - 92-112
OPTIMIZING SKIN CANCER DETECTION IN
THE USA HEALTHCARE SYSTEM USING DEEP
LEARNING AND CNNS
Md Nasiruddin
Department of Management Science and Quantitative Methods, Gannon
University, Erie, PA, USA
Mohammad Abir Hider
Master of Science in Business Analytics, Grand Canyon University, Phoenix,
AZ, USA
Rabeya Akter
Master of Science in information technology. Washington University of
Science and Technology, Alexandria, VA, USA
Shah Alam
Master of Science in Information Technology, Washington University of
Science and Technology, Alexandria, VA, USA.
MD Rashed Mohaimin
MBA in Business Analytics, Gannon University, Erie, PA, USA
MD Tushar Khan
Master of Science in Business Analytics, Trine University, Angola, IN, USA
Abdullah AL Sayeed
Master of Business Administration in Project Management, Central
Michigan University, Mt Pleasant, MI, USA
Afrin hoque jui
Management Sciences and Quantitative Methods, Gannon University, Eria,
PA, USA
RESEARCH ARTICLE
Open Access
THE USA JOURNALS
THE AMERICAN JOURNAL OF MEDICAL SCIENCES AND PHARMACEUTICAL RESEARCH
(ISSN
–
2689-1026)
VOLUME 06 ISSUE12
93
https://www.theamericanjournals.com/index.php/tajmspr
INTRODUCTION
Background and Motivation
Skin cancer is the most common cancer in the USA,
with millions of new cases reported each year. The
two main types of skin cancer include aggressive,
life-threatening melanoma and less lethal, though
potentially very morbid if left unattended, non-
melanoma types: basal cell carcinoma and
squamous cell carcinoma. According to the
American Cancer Society, that means an over 99%
five-year survival rate for early-stage melanoma,
against just 27% in the case of late-stage detection
(Rahman et al., 2023). This further signifies the all-
important question of diagnosis in time and with
high accuracy. Despite its importance, early skin
cancer detection faces various systemic barriers in
the US. These range from a shortage of
dermatologists and limited access to care in rural
and underserved areas to high costs related to
diagnostic procedures. Moreover, diagnostic
accuracy significantly varies among health
professionals; some depend on subjective visual
examination, which is prone to human error.
Advanced imaging techniques such as dermoscopy
have made the diagnosis more accurate, but their
effectiveness greatly relies on the clinician's
expertise (Saleh et al., 2023).
Al Amin et al. (2023), states that recent
developments in AI, especially deep learning, give
promising solutions to these challenges. Deep
learning is a subset of artificial intelligence; a class
of such algorithms called convolutional neural
networks has achieved great success in image
recognition tasks, medical imaging included. By
learning patterns in large datasets, CNNs can
classify skin lesions with accuracy comparable to,
or sometimes exceeding, that of dermatologists.
These AI-driven solutions, when integrated into
the USA health system, might change the face of
early diagnosis, reduce disparity in healthcare, and
Abstract
THE USA JOURNALS
THE AMERICAN JOURNAL OF MEDICAL SCIENCES AND PHARMACEUTICAL RESEARCH
(ISSN
–
2689-1026)
VOLUME 06 ISSUE12
94
https://www.theamericanjournals.com/index.php/tajmspr
result in improved patient outcomes (Bowmik et
al., 2023; Dutta et al., 2024).
Objectives
The principal aim of this research project is to
devise, curate, and propose a deep-learning CNN
methodology for skin cancer detection in the USA.
The specific objectives of this research are: To
apply deep learning methodologies for the
detection of skin cancer using dermoscopic images,
maintaining sensitivity and specificity at high
values. To develop a model of a convolutional
neural network adapted to classify the different
types of skin cancers, especially melanoma, BCC,
and SCC. This includes the performance evaluation
of the model on standard metrics of accuracy,
precision, recall, and F1-score, with further
consideration regarding clinical applications
within the US healthcare system.
LITERATURE REVIEW
According to Islam et al. (2023), Skin Cancer and
its Detection in the USA Skin cancer is the most
diagnosed cancer in the United States, with
millions of individuals receiving a skin cancer
diagnosis every year. The main types of skin cancer
are basal cell carcinoma, squamous cell carcinoma,
and melanoma. BCC and SCC are usually referred
to as non-melanoma skin cancers and are usually
highly curable if they are left early. Melanoma,
though rare, becomes much more dangerous
because it has great tendencies for metastasis. The
American Cancer Society estimates that in the USA,
melanoma is the cause of more deaths than any
other form of skin cancer, with 97,610 new cases
and 7,990 deaths projected for 2023(Hossain et al.
2024; Hider et al. 2024). The high prevalence of
skin cancers creates a huge burden on the patients,
as well as on the healthcare system. Direct costs
entail diagnosis, treatments, and follow-up over a
long period. Indirect costs are manifested as lost
workdays along with a reduction in general quality
of life. Accordingly, public health campaigns
emphasizing prevention and early detection raise
the chances of prognosis being exceedingly better.
However, at every turn, disparities in dermatologic
care add to the stress-especially in rural versus
underserved areas (Ghosh et al., 2024).
Current Diagnostic Techniques of Skin and
Their Limitations
1. Visual Inspection by Clinicians
As per Jaber & Akbas (2024), visual inspection
remains the first line of defense in diagnosing skin
cancer. Clinicians evaluate lesions based on their
asymmetry, border irregularity, color variation,
diameter, and evolution using the ABCDE criteria.
Though easy and cheap, it is highly dependent on a
clinician's experience and expertise. Studies
demonstrate big variability in diagnostic skills,
particularly among non-specialist providers. This
subjectivism often may lead to the under-sighting
of a suspicious case or unnecessary biopsy.
2. Dermatoscopy
Dermoscopy is a non-invasive method that allows
the professional, using polarized light and
magnification, to visualize structures subsurface in
the skin that may go unrecognized by the naked
eye. Thus, this technique is considered one of the
finest methods for diagnosing melanoma and
distinguishing malignancies from benign lesions
(Musthafa et al., 2023). Still, dermoscopy itself
does require extensive training and can be highly
related to an operator's learning curve concerning
diagnostic accuracy. Being only available to a very
restricted number of clinicians, in turn, reduces
equitable access to care and even further
exacerbates cancer outcome disparities in many
diverse geographical regions.
3. Histopathology
Histopathology is usually the gold standard in the
diagnosis of skin cancers, where microscopic
THE USA JOURNALS
THE AMERICAN JOURNAL OF MEDICAL SCIENCES AND PHARMACEUTICAL RESEARCH
(ISSN
–
2689-1026)
VOLUME 06 ISSUE12
95
https://www.theamericanjournals.com/index.php/tajmspr
investigations of biopsied tissue are performed. It
allows for the final confirmation of malignancy and
typing of cancer. However, this technique is
invasive, time-consuming, and expensive. There is
always a delay for the patients while waiting for
biopsy results, and more than 80% of performed
biopsies are unnecessary, which indicates the need
to improve the pre-biopsy diagnostic tool(Nancy et
al., 2023).
4. Reflectance Confocal Microscopy (RCM)
According to Nasiruddin et al. (2024), RCM is a
highly advanced imaging technique that enables
one to visualize the skin, at a cellular level and in
real-time, in high resolution. It allows physicians to
diagnose skin lesions without a biopsy. While RCM
decreases the need to perform invasive
procedures, it is costly and requires special
equipment; not only that, it puts a high demand on
skilled personnel for image interpretation, so it is
not commonly made available in routine clinical
practice. 5. Computer-Aided Diagnosis Systems
These CAD systems analyze dermoscopic images
by using algorithms and then make diagnostic
recommendations to clinicians. While the use of
such CAD systems enhances diagnostic accuracy
and reduces human error, their performance still
largely depends on the quality of the training
datasets. Most CAD systems are also not integrated
into routine clinical workflows and, in practice,
remain confined to a few well-equipped health
centers (Lilhore et al., 2023).
Deep Learning in Medical Imaging
Deep learning is a subclass of artificial intelligence;
it involves neural networks that are multilayered
and are trained to find patterns and features in
data. As opposed to traditional machine learning,
where explicit feature extraction is necessary,
deep learning learns the relevant features directly
from the training data. It has made it more apt for
the analysis of complex datasets such as medical
images. Deep learning, has shown phenomenal
promise in diagnosis across many medical
domains-radiology,
pathology,
and
dermatology(Sha et al., 2023). Such detection
ranges from the development of abnormalities
within chest X-rays to tumor classification in
histopathological slides and assessment of diabetic
retinopathy from retinal scans. It has thereby
positioned deep learning as probably one of the
most transformational tools in modern medicine
to
process
high-dimensional
data
with
unparalleled precision.
Empirical studies conducted by Zareen et al.,
(2024), deploying CNNs for the detection of skin
cancer have presented tremendous improvements
in diagnostic accuracy and efficiency. These utilize
deep learning techniques in the analysis of
dermatoscopic images that allow for the early
identification of skin lesions that may be
malignant. Among them, one of the notable works
proposed by Obayya et al. (2024), an optimized
CNN architecture that enhanced skin cancer
diagnosis using a very rich dataset called the
HAM10000, comprising dermatoscopic images. In
the proposed design, they developed a
sophisticated
model
comprising
several
convolutional, pooling, and dense layers to capture
complex visual features in skin lesions. This
approach included some interesting data
augmentation strategies to handle the class
imbalance in the dataset, enabling higher
diagnostic precision that could democratize
dermatology care, especially where specialist
expertise and/or access are limited.
The second major work by Saleh et al. (2023),
proposed the "Light-Dermo" model, which can be
seen as a light version of the CNN while
considering optimization towards real-time
applications. The given model used the mechanism
of channel-wise attention with integrated Squeeze-
and-Excitation blocks for improvements in the
THE USA JOURNALS
THE AMERICAN JOURNAL OF MEDICAL SCIENCES AND PHARMACEUTICAL RESEARCH
(ISSN
–
2689-1026)
VOLUME 06 ISSUE12
96
https://www.theamericanjournals.com/index.php/tajmspr
classifying accuracy along with computation
efficiency. On that account, the given research
indicated an improved performance using these
state-of-the-art models in previous works, for
example, 93.16% training accuracy and 91.93%
test accuracy among seven classes for PSLs2. This
highlights the potential for accuracy and practical
use in the clinical setting that CNNs have. A new
deep CNN approach was then proposed to address
class imbalance problems inherent in skin cancer
datasets. This model enhanced not only the
accuracy of classification but also showed
resilience in various challenges related to the
detection of skin cancer3. Advanced techniques,
including
transfer
learning
and
data
preprocessing, were integrated into CNNs to make
them more capable of distinguishing between
benign and malignant lesions.
Convolutional Neural Networks
According to Nancy et al., (2023), CNNs are a
special form of neural networks designed to
operate on grid-like data, such as images.
Architecture: Multiple layers, each performing a
specialized function:
Convolution Layers: These are the layers through
which filters are applied to the input image to
extract features that are related to edges, textures,
and patterns.
Activation Functions: Non-linear functions such as
ReLU introduce non-linearity, thereby helping the
network learn even the most complex
relationships.
Pooling Layers: Layers that reduce the spatial
dimensions of feature maps, hence emphasizing
important information and reducing computation.
Fully Connected Layers: This is the last layer; it
summarizes the features in the convolutional
layers to classify them.
Softmax Layer: The probability over each class is
attained from this layer, and the model uses that
for labeling an input image.
The hierarchical architecture of CNNs facilitates
them to learn low-level features (e.g., edges) in
initial layers and higher-level features (e.g., lesion
shapes or patterns) in deeper layers. This
capability also makes CNNs particularly effective
in tasks such as skin cancer detection, where
identifying subtle visual differences is key.
Key Achievements and Applications of CNNs in
Healthcare
Due to the improvements in model architecture
and computation resources, there has been fast
development of CNNs in healthcare. The main
achievements include:
Transfer Learning: Pre-trained models, such as
ResNet, VGGNet, and Efficient-Net, have been fine-
tuned for medical imaging tasks, reducing the need
for extensive training data.
Attention Mechanisms: These include various
techniques that allow models to concentrate on
specific parts of the image when necessary; this
helps raise the accuracy in specific tasks such as
lesion detection.
Explain-ability Tools: Techniques like Grad-CAM
give visual explanations for model predictions,
hence enhancing interpretability and clinical trust.
Successful modern deployments of CNNs in
healthcare include:
Radiology: CNN is used for detecting several
critical features, which include lung nodules and
other fractures and tumors based on various types
from the given input imaging modality obtained
normally by X-rays or CT scans.
Pathology: Whatever the case, the CNNs have
certainly brought a revolution in the analysis of
histopathology images for identifying cancerous
cells.
THE USA JOURNALS
THE AMERICAN JOURNAL OF MEDICAL SCIENCES AND PHARMACEUTICAL RESEARCH
(ISSN
–
2689-1026)
VOLUME 06 ISSUE12
97
https://www.theamericanjournals.com/index.php/tajmspr
Ophthalmology: The use of automation in
detecting diabetic retinopathy and age-related
macular degeneration has enhanced the efficiency
of its screening.
DATA COLLECTION AND PREPROCESSING
Data Sources
The dataset for the current research project was
retrieved from the Kaggle website, particularly,
The ISIC 2016 Skin Cancer Dataset contained
dermoscopic images that were used for skin cancer
classification. In this dataset, there were 1271
images of two classes of skin cancer, namely
Malignant and Benign. These images were then
gathered from the ISIC archive. The dataset was
then divided into a training set consisting of 1022
images and a test set consisting of 249 images
(Zihad, 2023). This dataset was used for training
and testing machine learning models for skin
cancer classification. The dataset is also useful in
the development of new image-processing
techniques for skin cancer detection and diagnosis.
In light of working with skin cancer data sets, the
most important ethical concerns refer to
compliance with data privacy legislation to protect
patient rights and develop trust in AI-driven health
solutions. For instance, in the United States, data
should be covered under the Health Insurance
Portability and Accountability Act, called HIPAA. It
demands vigorous protection for Protected Health
Information in terms of security. For example, data
de-identification requires removing all identifiable
information related to patients; a person's name,
date of birth, and also medical record numbers all
represent sources of personally identifiable
information in healthcare records. In this, some of
the key ethics that we considered include informed
consent regarding the use of a person's data,
insight into where this data is going to be used, and
anti-bias resulting in equitable performance of
their models, irrespective of different skin types or
geographical location.
Data Preprocessing
Effective data preprocessing is critical in making
any skin cancer dataset suitable for the training of
machine learning models. The first step was
cleaning, where images were reviewed to
eliminate duplicate samples, mislabeled samples,
and corrupted files, ensuring the dataset was
representative of the target classes. Secondly,
**image normalization** is performed for scaling
pixel values, usually within a range between 0 and
1. This normalizes the input and speeds up the
convergence of the model during training. Thirdly
**resizing** of all images to a fixed dimension, such
as 224x224 pixels, simply makes them compatible
with typical CNN architectures such as ResNet or
VGGNet since they require a fixed input size. To
increase diversity in the dataset and prevent
overfitting, some augmentation techniques are
adopted, including rotation, flipping, zooming, and
brightness adjustment. These augmentations
simulate real-world variations, enabling the model
to generalize better across unseen data. Fourth, the
dataset was split into training, validation, and test
sets, often using stratified sampling to maintain
class balance. The described preprocessing
pipeline is complete in ensuring that the dataset is
clean, standardized, and representative of a
phenomenon, which forms a good basis for robust
model training.
Exploratory Data Analysis (EDA)
THE USA JOURNALS
THE AMERICAN JOURNAL OF MEDICAL SCIENCES AND PHARMACEUTICAL RESEARCH
(ISSN
–
2689-1026)
VOLUME 06 ISSUE12
98
https://www.theamericanjournals.com/index.php/tajmspr
Figure 1: Exhibits Class Distribution in the Skin Cancer Dataset
The histogram above showcases a high-class imbalance in the Skin Cancer dataset, dominated by the class
"nv"-most probably with about 7,000 samples, while all other classes have much fewer representatives.
The second most represented classes, "mel" for melanoma and "bk" for benign keratosis-like lesions, have
about 1,000
–
1,200 samples each, showing a sharp drop from the dominant class. This is followed by "bcc"
standing for basal cell carcinoma, which has 600
–
800 samples, while "akiec" (Actinic keratosis) has less
than 600. The rare classes include "vasc"(vascular lesions) and "df" (dermatofibroma), representing
vascular lesions and dermatofibroma respectively, each having less than 200 samples. This is a challenge
in the training of machine learning models since the dominance of the class "nv" might result in biased
predictions and reduce the capability of the model to detect classes that are less frequent yet clinically
critical, such as melanoma. To handle the class imbalance, data augmentation, resampling techniques, or
class-weighted loss functions were necessary to make sure the model performance was robust and fair.
Figure 2: The Type of Medical Diagnostics Performed
THE USA JOURNALS
THE AMERICAN JOURNAL OF MEDICAL SCIENCES AND PHARMACEUTICAL RESEARCH
(ISSN
–
2689-1026)
VOLUME 06 ISSUE12
99
https://www.theamericanjournals.com/index.php/tajmspr
This histogram represents the distribution of different medical diagnostic techniques from a dataset.
"histo" probably is the abbreviation for histopathology, which was over 5,000 and therefore the most
popular diagnostic technique in the dataset. "Follow_up" follows with a count of roughly 3,500, further
elaborating that it is also a key technique used mostly for follow-up or tracking certain skin conditions.
The less frequently used method, "consensus," ranges at about 1,000 counts, while the least used is
"confocal" diagnostics, with less than 100 counts. The above distribution indicates reliance on
histopathology as the gold standard for diagnosis, while other methods, such as follow-ups, are
supportive. The very minimal utilization of confocal diagnostics perhaps reflects its specialized nature,
its cost, or its availability. This distribution underlines the interest in investigating complementary
methods to diversify and possibly streamline diagnostic workflows.
Figure 3: Displays the Body Parts Most Susceptible for Skin Cancer
This chart represents the distribution of skin cancer cases among different div parts and pinpoints
those parts that are most prone to this disease. The highest number of cases is recorded on the back, with
more than 2,000 cases, followed by the lower extremity and trunk, having counts above 1,500 and 1,000,
respectively. The upper extremity and abdomen are also well represented, each having about 800
–
1,000
cases. There is a moderate count for the face, chest, and foot between 500 and 700; whereas for the neck,
scalp, hand, and ear, there are less than 300 cases noted. Rare sites include the genital region and acral
sites, with very minimal counts. The "unknown" category indicates a small but notable portion of cases
where the location was unspecified. The pattern of distribution underlines that the focused screening has
to be performed on high-risk areas, especially the back and lower extremities, considering, however,
comprehensive coverage due to less common sites.
THE USA JOURNALS
THE AMERICAN JOURNAL OF MEDICAL SCIENCES AND PHARMACEUTICAL RESEARCH
(ISSN
–
2689-1026)
VOLUME 06 ISSUE12
100
https://www.theamericanjournals.com/index.php/tajmspr
Figure 4: Portrays Distribution of Age
This histogram above reflects the distribution of age in a dataset. The distribution appears to be right-
skewed, peaking in the 40-45 age group. This means that the majority of the people who form part of this
dataset are middle-aged. With increasing age, the count goes down very gradually to show a smaller
proportion of older ones. The long tail to the right is extended more, showing again that this distribution
is right-skewed. The overall histogram represents the distribution of age in the dataset, where middle-
age brackets are more populated.
Figure 5: Visualizes Gender Distribution
This histogram portrays the distribution graph of gender in a dataset: the majority of constituent
members of the dataset are males, about 5,500. The number is around 4,500 for females, which means
there is an imbalance in keeping up the gender level. The third category, "unknown", corresponds to a
very small figure compared to the other two sections, meaning that most sexes are known in this lot.
Overall, the shown histogram reflects a clear male-gender dominance within the dataset or population.
THE USA JOURNALS
THE AMERICAN JOURNAL OF MEDICAL SCIENCES AND PHARMACEUTICAL RESEARCH
(ISSN
–
2689-1026)
VOLUME 06 ISSUE12
101
https://www.theamericanjournals.com/index.php/tajmspr
Figure 6: Exhibits Dermoscopic Images with Different Kinds of Skin Lesions
Above are dermoscopic images, corresponding to different kinds of skin lesions. Each image is associated
with its diagnosis: akiec is the label used for actinic keratoses, bcc means basal cell carcinoma, bkl stands
for benign keratosis, mel refers to melanoma, nv stands for melanocytic nevus, and vascular lesions are
defined as vasc. Different colors, texture patterns, and shapes of lesions represent their wide variability.
While some lesions appear as elevated nodules or plaques, others may be flat or ulcerated. The colors
range from light brown to various tones of dark brown and black; some lesions contain shades of red or
blue. The great variability in the appearance of skin cancer accentuates the challenge in diagnosis, thus
placing considerable demands on methods for distinguishing benign from malignant lesions.
Figure 7: Depicts the Skin Cancer Dimensions, Aspect Ratio, and File Size Distribution
As displayed in the first histogram, it is uniformly distributed, with the majority of the images having a
width of about 600 pixels and a height of about 460 pixels. The second histogram shows that this aspect
THE USA JOURNALS
THE AMERICAN JOURNAL OF MEDICAL SCIENCES AND PHARMACEUTICAL RESEARCH
(ISSN
–
2689-1026)
VOLUME 06 ISSUE12
102
https://www.theamericanjournals.com/index.php/tajmspr
ratio is highly positively skewed, with most images having an aspect ratio of about 1.3. The third
histogram represents the file size distribution of the images; it appears to be right-skewed. Thus, most of
the images hold a file size between 100 and 200 KB. However, some may get up to over 300 KB. These
are broad indications that the size and shape of the majority of images are quite consistent but varying in
their file size presumably owing to quality and compression parameters of the camera or otherwise.
Figure 8: Showcases Skin Cancer Dataset-Mean Intensity Distribution
The histogram below shows the distribution of mean intensity values across the skin cancer dataset. This
distribution is approximately normal, peaking at around 160 and spreading from about 75 to 225. This
would suggest that most of the images in this collection have a mean intensity in the middle range, with
fewer with very low or very high mean intensities. Due to the normal shape, one would conclude that the
variation across the dataset is relatively stable in terms of mean intensity value.
Figure 9: Illustrates the Distribution of Pixel Intensities
THE USA JOURNALS
THE AMERICAN JOURNAL OF MEDICAL SCIENCES AND PHARMACEUTICAL RESEARCH
(ISSN
–
2689-1026)
VOLUME 06 ISSUE12
103
https://www.theamericanjournals.com/index.php/tajmspr
The above figure depicts the distributions of pixel intensities in red, green, and blue channels each for
vascular (vasc), melanocytic nevus (nv), melanoma (mel), dermato-fibroma (df), benign keratosis (bkl),
basal cell carcinoma (bcc), and actinic keratosis (akiec). These enable us to get an idea regarding the color
characteristics of different kinds of skin lesions. For example, the melanoma and dermato-fibroma
histograms are shifted toward the higher intensity values in the red channel, indicating a reddish color.
In contrast, benign keratosis and basal cell carcinoma have more concentration in the lower intensity
values in the red channel, indicating a brown or gray color. These variations in color distribution can help
distinguish the type of skin lesions and are possibly useful in automated classification systems.
Figure 10: Illustrates the Distribution of Skin Cancer- pair plot of image properties
Above is a pairs plot. A pairs plot is useful in visualizing the relationship between several different
properties of images within the Skin Cancer dataset. For the images, the diagonal plots give an idea of the
distribution for each separate property; these are: right-skewed for the means of intensity and the size of
the file while the highly-aspected ratio is strongly focused around 1.3 while the off-diagonal plots
represent a scatter plot of one property versus another. We observed weak positive correlations: average
intensity vs. file size as well as average intensity, and standard deviation of the intensities of the file. Weak
anticorrelation in aspect ratio about file size. By and large, though it should be that most associations
there in fact appear to be of real tenuosity, or else non-existent, as against the assumption that this
ensemble of image properties all will be highly interlinked -.
THE USA JOURNALS
THE AMERICAN JOURNAL OF MEDICAL SCIENCES AND PHARMACEUTICAL RESEARCH
(ISSN
–
2689-1026)
VOLUME 06 ISSUE12
104
https://www.theamericanjournals.com/index.php/tajmspr
METHODOLOGY
Model Architecture
The CNN proposed for this work is a deep-learning
architecture designed to address skin cancer
detection through dermoscopic images. The model
follows a sequential architecture with multiple
layers dedicated to the extraction of hierarchical
features from input images. The architecture
begins with an input layer where preprocessed
images are fed into the network, usually resized to
a fixed dimension, such as 224x224 pixels. After
that, several convolutional layers are applied to
identify the local features automatically, such as
edges, textures, and patterns, which form an
essential part of identifying unique features of skin
lesions. These convolutional layers are then
followed by a nonlinear actuation function like
ReLU, which would introduce non-linearity, hence
guaranteeing the network learns complex
patterns. The max-pooling layers serve to down-
sample the feature maps by convolution, hence
reducing its spatial dimensionality and subsequent
computational load, while retaining the most
salient features. To avoid overfitting, several
dropout layers technique that randomly
deactivates part of the neurons during training
into place, promoting better generalization
capability in the trained model. These feature
maps are then fed into fully connected layers that
combine all the learned features for final
classification. The output layer consists of a soft-
max activation function, which gives probability
scores for each class, such as benign, malignant,
and so on, to facilitate multi-class classification of
skin conditions.
This chosen architecture for CNN is based on an
appropriate theoretical rationale: it inherently
learns and selects spatial patterns from the
dermoscopic image in a hierarchical manner, to
build a decision boundary for the fine-grained
diagnosis of skin lesions. In this sense,
convolutional layers are best for extracting the
relevant spatial features, ideal in image analysis
tasks. The max-pooling layers contribute to
reducing the dimensionality such that the network
is efficient, and running while retaining the most
relevant features.
Besides, dropout was
implemented to avoid overfitting, which is an
important point when medical image datasets are
used since much variability in image quality and
characteristics of lesions may easily result in
overfitting. This last softmax layer is important for
multi-class classification, outputting a probability
distribution that indicates the likelihood of a given
skin lesion belonging to a specific class. The overall
architecture strikes a good balance between
feature extraction, efficiency, and robustness,
hence suitable for real-world clinical skin cancer
detection.
Training and Testing Framework
The skin cancer detection dataset was divided, into
clear-cut subsets for the fair assessment of the
performance of the model, in the training,
validation, and testing sets. This division was done
in a 70-20-10 split for devoting 70% of the data to
training, 20% going to validation, and testing
getting 10%. Testing data is used to allow the CNN
model to learn the different features across the
dermoscopic images. The validation set was
applied to monitor the performance of a model
iteratively during the training, where changes may
be made to avoid over-fitting. The test set was used
only for the very final evaluation of the
generalization capability of the model, Fairly
giving an unbiased estimate of the performance.
Finally, to add more robustness, k-fold cross-
validation was done. This technique has some
variations in which the dataset is divided into k
subsets, and then multiple trainings of the model
are done, each time using one subset as the
validation set while the rest of the data are used for
THE USA JOURNALS
THE AMERICAN JOURNAL OF MEDICAL SCIENCES AND PHARMACEUTICAL RESEARCH
(ISSN
–
2689-1026)
VOLUME 06 ISSUE12
105
https://www.theamericanjournals.com/index.php/tajmspr
training. Cross-validation helps ensure that a
model is not biased toward a particular subset of
the data and may provide a more general estimate
of its performance across the entire dataset.
Hyperparameter Tuning
The optimization of hyperparameters is very
crucial to optimize the performance of the model.
The three major hyperparameters of a CNN are
learning rate, batch size, and the number of epochs.
The learning rate controls the rate at which
changes in the weights of the model take place
during training, which impacts the speed and
stability of convergence. The batch size defines the
number of samples that will be processed before
updating the weights of the model. This value is a
trade-off that influences the efficiency and
memory usage during training. The number of
epochs is the number of times a model iterates
over the whole dataset. Now, for this, there is a set
of hyperparameters that one often optimizes using
techniques such as grid search and random search.
Grid search exhaustively checks all possible
combinations
of
hyperparameters
within
predefined ranges and assures that the best set is
chosen. On the other hand, random search
randomly selects hyperparameters in a certain
range and is often considerably faster, particularly
when large search spaces are considered.
Moreover, this process may be automated by tools
like Keras Tuner or Optuna, which effectively
explore
the
hyperparameter
space
and
recommend the best configuration for the best
performance.
Performance Evaluation Metrics
To assess the performance of the CNN algorithm
for skin cancer detection, several proven metrics
are utilized, namely, accuracy, precision, recall, and
F1-Score. Accuracy provides an overall idea of the
percent of right predictions-a percept of overall
general performance. Nevertheless, since skin
cancer image datasets used are highly imbalanced
in this context, it would not be sensible or practical
to rely only upon these measures. Therefore,
alongside
accuracy,
other
important
measurements considered are precision, recall,
and F1 score. Precision quantifies how many of
those positive predictions-e.g., true cases of
melanoma-are indeed correct, a critical measure to
avoid false positives in medical diagnostics. Recall,
on the other hand, is a metric that measures the
capability of the model to identify all actual
positive instances, such that no malignant cases go
unnoticed. The F1-score is the harmonic mean of
precision and recall; hence, it provides a balanced
measure of the model's ability to correctly identify
positives and minimize false negatives. Finally,
ROC-AUC, or Receiver Operating Characteristic -
Area Under the Curve, will be used to assess the
model's discriminatory ability across all
thresholds where higher AUC means better overall
performance in classification. These are
complementary and provide a complete metric to
address how well the model did, in particular, to
handle what's usually imbalanced classes of very
rare presence of melanoma as compared to benign
lesions.
RESULTS
Model Performance
THE USA JOURNALS
THE AMERICAN JOURNAL OF MEDICAL SCIENCES AND PHARMACEUTICAL RESEARCH
(ISSN
–
2689-1026)
VOLUME 06 ISSUE12
106
https://www.theamericanjournals.com/index.php/tajmspr
Figure 11: Visualizes Training Validation & Accuracy
These two plots show the training and validation loss and accuracy curves of some machine learning
models throughout 6 epochs. The plot on the left shows the training loss trending downward linearly; the
validation loss is initially going down, reaches a minimum around epoch 3, and then starts to increase.
This means that by epoch 3, this model begins to overfit the training data. The chart on the right shows
smoothly increasing training accuracy and validation accuracy which follows suit until a maximum
around Epoch 3 before slowly starting its decline. Further evidence of said overfitting behavior
concerning loss curves was provided earlier. The best performance is achieved at epoch 3, where the
validation loss is minimum and the validation accuracy is maximum. These plots highlight the importance
of monitoring both training and validation metrics to prevent overfitting and identify the optimal number
of training epochs.
Table 1: Displays CNNs Classification Report
precision recall f1-score support
nv 0.97 0.98 0.97 1367
mel 0.96 0.96 0.96 1350
bkl 0.89 0.94 0.92 1317
bcc 0.97 1.00 0.98 1352
akiec 0.87 0.84 0.86 1401
vasc 0.99 1.00 1.00 1347
df 0.93 0.87 0.90 1414
accuracy 0.94 9548
macro avg 0.94 0.94 0.94 9548
weighted avg 0.94 0.94 0.94 9548
The table above depicts the classification report of the skin cancer detection model performance on this
dataset, containing eight classes of skin lesions, namely, nv (melanocytic nevus), mel (melanoma), bk
THE USA JOURNALS
THE AMERICAN JOURNAL OF MEDICAL SCIENCES AND PHARMACEUTICAL RESEARCH
(ISSN
–
2689-1026)
VOLUME 06 ISSUE12
107
https://www.theamericanjournals.com/index.php/tajmspr
(benign keratosis), bcc (basal cell carcinoma), akiec (actinic keratosis), vasc (vascular lesion), and df
(dermatofibroma). It could be realized from this report that the model obtained a very high precision,
recall, and F1 score over all classes, with a general accuracy of 94% for this multi-class problem. This
model was very good, both in precision since it correctly identifies the actual positive cases and in recall,
where it does not have false positives. Also, the macro and weighted average scores are very good in
general. However, the slight overperformance of classes such as bcc and vasc concerning others like bkl
and akiec could be an area of improvement.
Table 2: Exhibits CNNs Confusion Matrix
The confusion matrix offered a detailed
breakdown of the classification performance
across multiple classes, particularly showing the
true labels versus the predicted labels. It managed
to classify 1,338 instances of the class "
melanocytic nevus " correctly but mislabeled 22 as
"melanoma" and 7 as "benign keratosis." For the
class "melanaoma", the model had a true positive
count of 1,296 but misclassified 8 instances as
"melanocytic nevus" and 20 as "benign keratosis."
In this class, "df" also showed very good
performance with 1,231 correct predictions,
though there are still misclassifications, especially
in its instances identified as "actinic keratosis)"
with 96 and "benign Keratosis " with 13. In
summary, most classes have relatively high values
on the diagonal, but the misclassifications do raise
concerns that the network may need
improvement, particularly between similar classes
like "bkl" and "mel."
DISCUSSION
Clinical Implications
The developed proposed CNN model for skin
cancer detection has great potential to support
human clinical decision-making in all dermatology.
This developed model automates the various
analyses of dermoscopy images, hence acting as
just an adjunct tool for active dermatologists,
which shall enable fast and accurate skin lesion
assay. It is especially valued in those settings
where dermatologic overload or access to dermal
specialists is poor outside major cities in the USA,
The high precision and recall rates of the model can
reduce false positives and false negatives, hence
the early detection of malignant cases and
avoidance
of
unnecessary
biopsies
or
interventions on benign cases. Moreover, the
ability of CNNs to detect subtle patterns that may
be imperceptible to the human eye further
THE USA JOURNALS
THE AMERICAN JOURNAL OF MEDICAL SCIENCES AND PHARMACEUTICAL RESEARCH
(ISSN
–
2689-1026)
VOLUME 06 ISSUE12
108
https://www.theamericanjournals.com/index.php/tajmspr
enhances diagnostic accuracy. This means that
clinicians better-informed decisions, at least a
reduction in diagnostic uncertainty-something so
crucial in diseases like melanoma.
Results have shown that this CNN can easily be
integrated into diagnosis workflows in normal
dermatological practice to offer a second opinion
or even a pre-screening tool for dermatologists.
Teledermatology, just now gaining steam in the
USA and even more so post-COVID-19 pandemic,
may be greatly helped by such technology. This will
enable patients to upload photos of skin lesions
from the comfort of their homes; the CNN model
will scan them for flagging probable malignant
lesions for further evaluation. Moreover, the tool
assists in triaging cases, as it prioritizes patients
with high-risk lesions to immediate care. These
applications will not only raise the efficiency of
dermatological services but also ensure equal
access to quality care, especially for patients in
remote locations.
Integration into USA healthcare systems
These highlighted aspects of the CNN model could
provide transformative benefits in their
incorporation within healthcare systems in the
United States. Early diagnosis, mainly of skin
cancer, has improved survival rates due to
interventions at an early stage. The model can
ensure regularity in diagnosis and consistency to a
greater degree with diagnostic errors that are well-
known in dermatology based on subjective human
judgments. Furthermore, this model can be
incorporated into EHRs in a way that automatic
analysis of dermoscopic images at routine
checkups reduces the workload of dermatologists,
hence saving time to be spent on consulting
patients. This could make the health and medical
care provided cheaper since several biopsies and
treatments would not be carried out.
Nevertheless, the integration of such AI models as
CNNs into clinical workflows is not without its
challenges. One major consideration is to ensure
interoperability with existing healthcare systems,
such as EHR platforms used across hospitals and
clinics. The model needs to be validated across
diverse populations and different imaging devices
to ensure its generalizability and reliability in real-
world scenarios. Resistance to adopting AI tools
from clinicians, often driven by concerns about
trust, job security, or unfamiliarity with the
technology, must also be addressed. These barriers
can be minimized with extensive training
programs and well-articulated guidelines on the
effective use of AI tools. Besides, regulatory
compliance will have to be maintained with
organizations such as the FDA, and ethical
standards followed to nurture trust and
accountability in deploying AI models.
Limitations and Challenges
While the CNN model has a lot of advantages, some
ethical
and
technical
challenges
need
consideration for successful deployment. Some of
the ethical concerns involve using patient data to
train AI models, including privacy, consent, and
security. Compliance with regulations such as
HIPAA is very important in the USA, and
anonymization of datasets is required for the
protection of patient identity. Moreover, it is
important to have no datasets that induce bias, in
which changes in minority groups are
disproportionately affected, for a dataset to be
representative of the broader population. For
example, algorithms for skin cancer diagnosis
trained only on light-skinned populations have
performed very poorly on darker skin, underlining
the need to consider inclusive datasets.
Another limitation pertains to the quality and
interpretability of the model. CNNs are often
described as "black-box" systems, and their
decision-making processes are not always
THE USA JOURNALS
THE AMERICAN JOURNAL OF MEDICAL SCIENCES AND PHARMACEUTICAL RESEARCH
(ISSN
–
2689-1026)
VOLUME 06 ISSUE12
109
https://www.theamericanjournals.com/index.php/tajmspr
transparent. This lack of interpretability may pose
challenges
in
clinical
settings
where
understanding the rationale for a diagnosis is
important for patient care. Besides, changes in
image quality, lighting, and resolution across
datasets could affect model performance and raise
questions about its robustness in real-world
applications. Generalizability remains another
concern since models trained on one dataset may
not perform well on completely different datasets,
thus requiring extensive validation before clinical
deployment.
Future Research Directions
Overcoming pinpointed challenges by future
research will be vital to enhance the effectiveness
and applicability of the CNN model. The most
viable way to do this might be using larger datasets
that are more diverse, with images from people of
different
ethnicities,
age
brackets,
and
geographical regions. This would go a long way in
enhancing the model's generalization ability
across different patient populations and reducing
the risk of bias. The model could also apply
knowledge, with advances in transfer learning,
from pre-trained architectures, thus limiting the
need for large datasets and accelerating
development. Further, the model may be improved
by using complementary data from other
modalities, such as genetic information or patient
histories, which would enhance not only its
diagnostic performance but also yield a broader
view of the health status of a patient.
Another exciting area for exploration is real-time
skin cancer detection. Further work may,
therefore, be directed at an improved CNN model,
one that, through portable devices or using a
phone application, instantly provides analyses of
skin lesions. That will put patients in the position
to self-manage their status of skin health and take
timely medical advice whenever necessary.
Another very promising area is the integration of
AI models with personal treatment planning.
These CNNs may further develop the
characteristics of lesions and relate them to patient
outcomes to predict the most effective treatments
in any given case, advancing precision medicine in
dermatology. Ultimately, ongoing research and
collaboration between clinicians, researchers, and
technologists will be required to unlock the full
potential of CNN models to transform skin cancer
detection and treatment.
Impact on the USA Health Care System
Improved Diagnostic Precision and Early
Detection: With greatly enhanced accuracy and
timeliness, CNNs may bring a paradigm shift in the
diagnosis of skin cancer in the USA. The subtle
patterns and features that may be missed by the
human naked eye are picked up through
dermoscopic image analysis by the CNN model.
The resultant increased accuracy can lead to the
early detection of skin cancer, thereby allowing
timely intervention that will improve patient
outcomes.
Reduced Diagnostic Errors: One of the significant
benefits of AI-driven CNN models is identified in
the reduction of diagnostic errors. These models
will tend to provide consistent and objective
analyses that reduce the human error factor, which
might be caused by tiredness, inexperience, or an
element of subjective judgment. That would
increase reliability for a more perfect diagnosis
that can effectively be treated.
Efficiency
in
Clinical
Workflows:
The
incorporation of CNNs into the clinical workflow
may provide efficiency in diagnosis, reducing the
burden on dermatologists and other healthcare
professionals. With the automation of image
analysis, these models give a preliminary diagnosis
that may liberate clinicians to attend to more
complex cases and interpersonal interactions with
THE USA JOURNALS
THE AMERICAN JOURNAL OF MEDICAL SCIENCES AND PHARMACEUTICAL RESEARCH
(ISSN
–
2689-1026)
VOLUME 06 ISSUE12
110
https://www.theamericanjournals.com/index.php/tajmspr
patients. The resultant increase in efficiency can
then lead to shorter wait times, improved patient
satisfaction, and better use of healthcare
resources.
Cost-Effectiveness: Early detection and proper
diagnosis of skin cancer will lead to long-term cost
savings. The reason is that early detection and
treatment prevent expensive surgeries and
radiation therapies associated with the advanced
stages of skin cancer. It could also reduce burdens
on healthcare professionals and raise efficiency to
lower healthcare costs overall. Personalized
Treatment Plans AI-powered CNN models can help
in gleaning valuable insights into the patient-
specific characteristics that will allow the
development of personalized treatment plans.
Thus, these models can identify risk factors and
predict disease progression by analyzing the
patient's medical history, lifestyle factors, and
imaging data. The information may further be used
to tailor the appropriate treatment strategies by
recommending certain medications, therapies, or
surgical procedures for optimal outcomes for the
patients.
CONCLUSION
The principal aim of this research project is to
devise, curate, and propose a deep-learning CNN
methodology for skin cancer detection in the USA.
The dataset for the current research project was
retrieved from the Kaggle website, particularly,
The ISIC 2016 Skin Cancer Dataset contained
dermoscopic images that were used for skin cancer
classification. In this dataset, there were 1271
images of two classes of skin cancer, namely
Malignant and Benign. These images were then
gathered from the ISIC archive. The dataset was
then divided into a training set consisting of 1022
images and a test set consisting of 249 images. The
CNN proposed for this work is a deep-learning
architecture designed to address skin cancer
detection through dermoscopic images. The model
follows a sequential architecture with multiple
layers dedicated to the extraction of hierarchical
features from input images. To assess the
performance of the CNN algorithm for skin cancer
detection, several proven metrics are utilized,
namely, accuracy, precision, recall, and F1-Score.
The model obtained a very high precision, recall,
and F1-score over all classes, with a general
accuracy of 94% for this multi-class problem. This
model was very good, both in precision since it
correctly identifies the actual positive cases and in
the recall, where it does not have false positives.
The developed proposed CNN model for skin
cancer detection has great potential to support
human clinical decision-making in all dermatology.
This developed model automates the various
analyses of dermoscopy images, hence acting as
just an adjunct tool for active dermatologists,
which shall enable fast and accurate skin lesion
assay. Results have shown that this CNN can easily
be integrated into diagnosis workflows in normal
dermatological practice to offer a second opinion
or even a pre-screening tool for dermatologists.
REFERENCES
1.
Alam, S., Hider, M. A., Al Mukaddim, A., Anonna,
F. R., Hossain, M. S., khalilor Rahman, M., &
Nasiruddin, M. (2024). Machine Learning
Models for Predicting Thyroid Cancer
Recurrence: A Comparative Analysis.
2.
Al Amin, M., Liza, I. A., Hossain, S. F., Hasan, E.,
Haque, M. M., & Bortty, J. C. (2024). Predicting
and Monitoring Anxiety and Depression:
Advanced Machine Learning Techniques for
Mental Health Analysis. British Journal of
Nursing Studies, 4(2), 66-75.
3.
Bhowmik, P. K., Miah, M. N. I., Uddin, M. K.,
Sizan, M. M. H., Pant, L., Islam, M. R., & Gurung,
N. (2024). Advancing Heart Disease Prediction
through Machine Learning: Techniques and
THE USA JOURNALS
THE AMERICAN JOURNAL OF MEDICAL SCIENCES AND PHARMACEUTICAL RESEARCH
(ISSN
–
2689-1026)
VOLUME 06 ISSUE12
111
https://www.theamericanjournals.com/index.php/tajmspr
Insights for Improved Cardiovascular Health.
British Journal of Nursing Studies, 4(2), 35-50.
4.
Bortty, J. C., Bhowmik, P. K., Reza, S. A., Liza, I.
A., Miah, M. N. I., Chowdhury, M. S. R., & Al
Amin, M. (2024). Optimizing Lung Cancer Risk
Prediction with Advanced Machine Learning
Algorithms and Techniques. Journal of Medical
and Health Studies, 5(4), 35-48.
5.
Dutta, S., Sikder, R., Islam, M. R., Al Mukaddim,
A., Hider, M. A., & Nasiruddin, M. (2024).
Comparing the Effectiveness of Machine
Learning Algorithms in Early Chronic Kidney
Disease Detection. Journal of Computer Science
and Technology Studies, 6(4), 77-91.
6.
Ghosh, H., Rahat, I. S., Mohanty, S. N., Ravindra,
J. V. R., & Sobur, A. (2024). A Study on the
Application of Machine Learning and Deep
Learning Techniques for Skin Cancer
Detection. International Journal of Computer
and Systems Engineering, 18(1), 51-59.
7.
Hasan, E., Haque, M. M., Hossain, S. F., Al Amin,
M., Ahmed, S., Islam, M. A., ... & Akter, S. (2024).
CANCER DRUG SENSITIVITY THROUGH
GENOMIC DATA: INTEGRATING INSIGHTS
FOR PERSONALIZED MEDICINE IN THE USA
HEALTHCARE SYSTEM. The American Journal
of Medical Sciences and Pharmaceutical
Research, 6(12), 36-53.
8.
Hider, M. A., Nasiruddin, M., & Al Mukaddim, A.
(2024). Early Disease Detection through
Advanced Machine Learning Techniques: A
Comprehensive Analysis and Implementation
in Healthcare Systems. Revista de Inteligencia
Artificial en Medicina, 15(1), 1010-1042.
9.
Hossain, S., Miah, M. N. I., Rana, M. S., Hossain,
M. S., Bhowmik, P. K., & Rahman, M. K. (2024).
ANALYZING TRENDS AND DETERMINANTS
OF LEADING CAUSES OF DEATH IN THE USA:
A DATA-DRIVEN APPROACH. The American
Journal
of
Medical
Sciences
and
Pharmaceutical Research, 6(12), 54-71.
10.
Hossain, M. S., Rahman, M. K., & Dalim, H. M.
(2024). Leveraging AI for Real-Time
Monitoring and Prediction of Environmental
Health Hazards: Protecting Public Health in the
USA. Revista de Inteligencia Artificial en
Medicina, 15(1), 1117-1145.
11.
Islam, M. Z., Nasiruddin, M., Dutta, S., Sikder, R.,
Huda, C. B., & Islam, M. R. (2024). A
Comparative Assessment of Machine Learning
Algorithms for Detecting and Diagnosing
Breast Cancer. Journal of Computer Science
and Technology Studies, 6(2), 121-135.
12.
Jaber, N. J. F., & Akbas, A. (2024). Melanoma
skin cancer detection based on deep learning
methods and binary Harris Hawk optimization.
Multimedia Tools and Applications, 1-14.
13.
Lilhore, U. K., Simaiya, S., Sharma, Y. K., Kaswan,
K. S., Rao, K. B., Rao, V. M., ... & Alroobaea, R.
(2024). A precise model for skin cancer
diagnosis using hybrid U-Net and improved
MobileNet-V3
with
hyperparameters
optimization. Scientific Reports, 14.
14.
Musthafa, M. M., TR, M., V, V. K., & Guluwadi, S.
(2024). Enhanced skin cancer diagnosis using
optimized CNN architecture and checkpoints
for
automated
dermatological
lesion
classification. BMC Medical Imaging, 24(1),
201.
15.
Nancy, V. A. O., Prabhavathy, P., Arya, M. S., &
Ahamed, B. S. (2023). Comparative study and
analysis on skin cancer detection using
machine learning and deep learning
algorithms.
Multimedia
Tools
and
Applications, 82(29), 45913-45957.
16.
Nasiruddin, M., Dutta, S., Sikder, R., Islam, M. R.,
Mukaddim, A. A., & Hider, M. A. (2024).
Predicting Heart Failure Survival with Machine
THE USA JOURNALS
THE AMERICAN JOURNAL OF MEDICAL SCIENCES AND PHARMACEUTICAL RESEARCH
(ISSN
–
2689-1026)
VOLUME 06 ISSUE12
112
https://www.theamericanjournals.com/index.php/tajmspr
Learning: Assessing My Risk. Journal of
Computer Science and Technology Studies,
6(3), 42-55.
17.
Obayya, M., Arasi, M. A., Almalki, N. S., Alotaibi,
S. S., Al Sadig, M., & Sayed, A. (2023). Internet of
things-assisted smart skin cancer detection
using metaheuristics with deep learning
model. Cancers, 15(20), 5016.
18.
Pant, L., Al Mukaddim, A., Rahman, M. K.,
Sayeed, A. A., Hossain, M. S., Khan, M. T., &
Ahmed, A. (2024). Genomic predictors of drug
sensitivity in cancer: Integrating genomic data
for personalized medicine in the USA.
Computer Science & IT Research Journal,
5(12), 2682-2702.
19.
Rahman, A., Karmakar, M., & Debnath, P.
(2023). Predictive Analytics for Healthcare:
Improving Patient Outcomes in the US through
Machine Learning. Revista de Inteligencia
Artificial en Medicina, 14(1), 595-624.
20.
Saleh, N., Hassan, M. A., & Salaheldin, A. M.
(2024). Skin cancer classification based on an
optimized convolutional neural network and
multicriteria
decision-making.
Scientific
Reports, 14(1), 17323.
21.
Shah, A., Shah, M., Pandya, A., Sushra, R.,
Sushra, R., Mehta, M., ... & Patel, K. (2023). A
comprehensive study on skin cancer detection
using artificial neural network (ANN) and
convolutional neural network (CNN). Clinical
eHealth.
22.
Zareen, S. S., Sun, G., Kundi, M., Qadri, S. F., &
Qadri, S. (2024). Enhancing Skin Cancer
Diagnosis with Deep Learning: A Hybrid CNN-
RNN Approach. Computers, Materials &
Continua, 79(1).
23.
Zihad, F. (2023, October 17). Skin Cancer
Dataset
ISIC
2016.
Kaggle.
https://www.kaggle.com/datasets/mdforiduz
zamanzihad/skin-cancer-dataset-isic-2016
