SKIN CANCER DETECTION USING EFFICIENTNET

A Mini Project Submitted to the

Department of Electronics and Communication Engineering,
in partial fulfillment of the requirements for the degree of
Bachelor of Technology
in
Electronics and Communication Engineering

Course Code: 20EC6554

Course Title: Mini Project-1

By

A Ram Mohith Sk Shaman K Harshith

Roll No: 218W1A04D1 Roll No: 218W1A04I9 Roll No: 218W1A04G0

Supervised By
Mr. [Link] Raja,[Link],(Ph.D)
Assistant Professor

Department of Electronics and Communications Engineering

Velagapudi Ramakrishna Siddhartha Engineering College
Kanuru, Vijayawada - 520007
 Department of Electronics and Communications Engineering
Velagapudi Ramakrishna Siddhartha Engineering College
Kanuru, Vijayawada - 520007

CERTIFICATE

This is to certify that the work entitled ‘SKIN CANCER DETECTION USING Efficient-
Net’ authored by A Ram Mohith, Sk Shaman, and K Harshith, has been carried out under
my supervision. To the best of our knowledge, this work is original and has not been
submitted elsewhere for any diploma or degree.

Supervisor

Mr. [Link] Raja,[Link],(Ph.D.)

Assistant Professor
Department of Electronics and Communications Engineering
Velagapudi Ramakrishna Siddhartha Engineering College
Kanuru, Vijayawada - 520007

ii
 Department of Electronics and Communications Engineering
Velagapudi Ramakrishna Siddhartha Engineering College
Kanuru, Vijayawada - 520007

DECLARATION

The work entitled ‘SKIN CANCER DETECTION USING EfficientNet’ which has been
carried out in the Department of Electronics and Communications Engineering at Ve-
lagapudi Ramakrishna Siddhartha Engineering College, is original and conforms to the
regulations of this Institute.

We understand the Institute’s policy on plagiarism and declare that no part of this work
has been copied from other sources or previously submitted elsewhere for any degree or
diploma.

(A Ram Mohith) (Sk Shaman) (G Harshith)

StudentID: 218W1A04D1 Student ID: 218W1A04I9 StudentID: 218W1A04G0

iii
 Acknowledgment

We would like to express our sincere gratitude to the Department of Electronics and
Communication Engineering at Velagapudi Ramakrishna Siddhartha Engineering Col-
lege for providing the necessary resources and support throughout the course of this
project.

We extend our appreciation to Dr.A V Ratna Prasad, the Principal of the institution, for
his encouragement. His visionary leadership has been instrumental in fostering an en-
vironment conducive to academic excellence and innovation.

Special thanks to Dr. Dhulipalla Venkata Rao, Head of the Electronics and Communica-
tion Engineering Department, for his valuable insights and constant support. we are also
grateful to our project guide, [Link] Raja, for his expertise and mentorship, which
significantly contributed to the success of this endeavor.

iv
Table of Contents

Certificate ii

Declaration iii

Acknowledgment iv

Table of Contents v

List of Figures vii

Abstract viii

1 Introduction 1
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.1 CNN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Motivation and Inspiration . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.5 Thesis Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 Literature Review 6
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 CNN architectures for Image Classification . . . . . . . . . . . . . . . 7
2.2.1 Alexnet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2.2 Resnet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2.3 Densenet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2.4 Mobilenet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3 Related Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

v
 2.4 Problem Statements . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.5 Proposed Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

3 Proposed Approach 13
3.1 Inrtoduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2 Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.3 EffecientNet B0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.4 Breif about methodology . . . . . . . . . . . . . . . . . . . . . . . . . 16

4 Results and Discussion 17

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.2 Evaluation and Discussion . . . . . . . . . . . . . . . . . . . . . . . . 18

5 Conclusion and Future Work 21

5.1 Conclusion and Future work . . . . . . . . . . . . . . . . . . . . . . . 22

References 23

vi
List of Figures

3.1 Skin Lesion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

3.2 The architecture of EffecientNet b0 CNN . . . . . . . . . . . . . . . . 15

4.1 EffecientNet B0 training progress . . . . . . . . . . . . . . . . . . . . 19

4.2 Confusion matrix for the test results. . . . . . . . . . . . . . . . . . . . 20

vii
 Abstract

This abstract provides an overview of an ongoing project that focuses on harnessing

Convolutional Neural Networks (CNN) and other deep learning techniques for the early
detection of skin cancer. Skin cancer, arising from unrepaired DNA damage in skin
cells, presents a significant health concern, underscoring the critical need for early de-
tection strategies. The project explores a range of lesion parameters, including symme-
try, color, size, and shape, utilizing CNN to discern benign conditions from malignant
melanomas.

Conducting a systematic review, we analyze research papers from reputable journals to

investigate the efficacy of CNN-based models and other deep learning frameworks. The
ongoing work aims to synthesize and present the dynamic landscape of early detection
methodologies, emphasizing both the strengths and limitations of employing CNN in
skin cancer diagnosis.

This project focuses on the development of a skin cancer detection system using deep
learning techniques. The ISIC dataset was utilized, and after preprocessing, the dataset
was divided into two classes for binary classification. An EfficientNet B0 neural net-
work was selected as the base model and was fine-tuned on a specific dataset. The
trained model was then tested on a separate dataset, achieving a high level of accuracy
in detecting skin cancer. This project demonstrates the effectiveness of deep learning
models in medical image analysis, specifically in the detection of skin cancer, which
can potentially aid in early diagnosis and treatment.

Keywords: deep learning; efficientnet; machine learning; melanoma.

viii
 Chapter 1
Introduction
 Chapter 1 – Introduction

1.1 Introduction
Skin cancer, a prevalent and concerning form of cancer, manifests when the DNA within
skin cells undergoes damage or mutation. This can be attributed to various factors, with
excessive sun exposure being the primary contributor. Prolonged exposure to Ultra-
Violet (UV) radiation from the sun can disrupt the intricate instructions encoded in the
DNA of skin cells, initiating abnormal cell growth and ultimately leading to the devel-
opment of cancer. Although skin cancer can manifest on any part of the skin, it is most
frequently observed in sun-exposed areas such as the face, neck, arms, and hands.

This project report delves into the intricacies of skin cancer, exploring its types and un-
derlying causes. The skin comprises three fundamental layers—squamous cells, basal
cells, and melanocytes. The outermost layer is prone to the development of non-melanoma
cancers, including Basal Cell Carcinoma (BCC) and Squamous Cell Carcinoma (SCC).
On the other hand, Melanoma, originating in melanocytes responsible for pigment pro-
duction, represents the most perilous form of skin cancer due to its potential to rapidly
metastasize if not detected early.

Understanding the causes of skin cancer is crucial for effective prevention and early de-
tection. Factors such as prolonged exposure to UV radiation, family history, chemical
exposures, and a weakened immune system contribute to the development of skin can-
cer. UV radiation, whether from the sun or artificial sources like tanning beds, emerges
as a prominent risk factor, causing DNA damage that leads to the formation of cancer-
ous cells. Additionally, genetic predisposition, chemical exposures, and compromised
immune systems play integral roles in elevating the susceptibility to skin cancer.

This project endeavors to provide a comprehensive exploration of skin cancer, shedding

light on its various types, causes, and implications. By unraveling the intricacies of this
prevalent malignancy, the aim is to contribute valuable insights for better prevention,
early detection, and ultimately, improved outcomes for individuals at risk or affected by
skin cancer.

1.1.1 CNN
Convolutional Neural Networks (CNN) represent a class of deep learning algorithms
specifically designed for visual data processing tasks. They are particularly adept at
handling image-related tasks due to their ability to automatically learn hierarchical rep-

2
 Chapter 1 – Introduction

resentations from input data. CNNs consist of convolutional layers that apply filters to
input data, capturing features at different levels of abstraction, and pooling layers that
reduce the spatial dimensions of the learned features.

Usage in Skin Cancer Detection:

In the realm of dermatology and medical imaging, CNNs have emerged as powerful
tools for the early detection and classification of skin cancer. Leveraging the inherent
capabilities of CNNs in feature extraction and pattern recognition, these networks can
analyze dermatoscopic images and assist in identifying potential malignancies.

In conclusion, CNNs have demonstrated significant potential in revolutionizing the field

of skin cancer detection. Their ability to automatically learn and extract relevant fea-
tures from dermatoscopic images contributes to early diagnosis and improved patient
outcomes. As research in this area progresses, the integration of CNNs into clinical
practice holds promise for enhancing the accuracy and efficiency of skin cancer diagno-
sis.

1.2 Motivation and Inspiration

The pursuit of skin cancer detection through deep learning techniques is profoundly mo-
tivating and inspiring on multiple fronts. Beyond the technological allure of employing
cutting-edge artificial intelligence, this research carries substantial medical significance.
Early detection of skin cancer, facilitated by accurate and efficient deep learning mod-
els, holds the promise of vastly improving treatment outcomes and subsequently ele-
vating survival rates. The global prevalence of skin cancer emphasizes the potential
widespread impact of such advancements, particularly in regions with limited access
to specialized healthcare. This interdisciplinary field not only fosters collaboration be-
tween computer scientists and medical professionals but also contributes to the ongoing
evolution of medical imaging and diagnostics. Personal stories of individuals whose
lives have been positively affected by early detection serve as poignant reminders of the
real-world impact of this research. Moreover, by contributing to the development of ac-
cessible and scalable diagnostic tools, researchers in this field play a role in promoting
preventive healthcare practices. The intellectual stimulation of overcoming challenges
and the ethical responsibility associated with advancing healthcare through technology
further amplify the motivation behind endeavors in skin cancer detection using deep
learning techniques. In essence, this work is a convergence of technological innovation,

3
 Chapter 1 – Introduction

medical progress, and a commitment to improving lives on a global scale.

1.3 Objectives
:

* Achieve Early and Accurate Diagnosis: Prioritize timely detection of skin cancer
to facilitate early [Link] diagnostic accuracy beyond traditional methods
through the implementation of deep learning models.

* Enhance Diagnostic Precision: Classify skin lesions into benign and malignant cate-
gories with a high level of precision. Develop automated systems seamlessly integrated
into existing clinical workflows to streamline the diagnostic process.

* Ensure Robust Model Design: Design models that demonstrate robustness, accommo-
dating variations in skin types and environmental conditions. Consider diverse demo-
graphic factors to ensure the inclusivity and effectiveness of the models across different
patient populations.

* Improve Interpretability: Focus on improving the interpretability of deep learning

models to foster trust among healthcare professionals. Enhance transparency in model
decision-making processes to facilitate collaboration between AI systems and human
experts.

* Address Scalability Challenges: Develop scalable models capable of handling large

and diverse datasets from various healthcare settings. Ensure adaptability to different
clinical environments, catering to the scalability needs of diverse healthcare infrastruc-
tures.

1.4 Scope
The scope for the implementation of deep learning techniques in skin cancer detection is
vast and holds significant potential for positive impact. Firstly, these technologies have
the potential to revolutionize current diagnostic practices, offering more accurate and
timely identification of skin cancer lesions, leading to improved patient outcomes. The
scalability of these models allows for widespread adoption in diverse healthcare settings,
making advanced diagnostic tools more accessible globally, especially in regions with
limited healthcare resources.

4
 Chapter 1 – Introduction

Moreover, the integration of deep learning into existing clinical workflows can stream-
line the diagnostic process, allowing healthcare professionals to make quicker and more
informed decisions. The automation of skin cancer detection not only increases effi-
ciency but also reduces the burden on healthcare systems.

The continuous evolution of deep learning models in this field ensures that ongoing re-
search can address emerging challenges and refine existing methodologies. Overall, the
scope for skin cancer detection using deep learning extends beyond technological ad-
vancements to encompass improvements in patient care, accessibility, and global public
health outcomes.

1.5 Thesis Organization

This thesis is structured into five chapters, each with a specific focus. A brief summary
of these chapters is presented below.

Chapter 1 serves as an introduction to this thesis, highlighting its main objective, moti-
vation, research problems, and contributions.

Chapter 2 provides an overview of important concepts and background knowledge rele-

vant to the domain. It also explores existing studies related to transliteration, highlight-
ing their limitations.

Chapter 3 outlines the complete methodology proposed in this thesis. It covers vari-
ous aspects such as dataset description, preprocessing techniques, feature encoding and
selection methods, performance measures, and the models employed.

Chapter 4 delves into the conducted experiments and presents detailed findings. It crit-
ically evaluates and discusses the outcomes of these experiments.

Chapter 5 concludes the work presented in this thesis and outlines potential directions
for future research.

5
 Chapter 2
Literature Review
 Chapter 2 – Literature Review

2.1 Introduction
The literature review for the project on skin cancer detection using deep learning tech-
niques serves as a comprehensive exploration of existing research, methodologies, and
advancements in the realm of automated diagnosis. With the increasing incidence of
skin cancer and the imperative need for early and accurate detection, this literature re-
view aims to provide a nuanced understanding of the landscape surrounding the appli-
cation of deep learning in skin cancer diagnosis.

The objectives of employing deep learning techniques in skin cancer detection are multi-
faceted, ranging from achieving early and precise diagnoses to ensuring the ethical and
real-world applicability of the developed models. In the following sections, we delve
into a synthesis of scholarly works, research papers, and developments in the field, ex-
amining the evolution of deep learning methodologies tailored for dermatological appli-
cations.

This review will explore the foundations of deep learning, focusing on Convolutional
Neural Networks (CNNs) and other relevant architectures that have shown promise in
image analysis and medical diagnostics. The emphasis lies on understanding the unique
challenges posed by skin cancer diagnosis, including the diversity of skin types, the
complexity of lesion patterns, and the need for interpretability in clinical settings.

2.2 CNN architectures for Image Classification

This section overviews the most commonly used CNN architectures in the image clas-
sification task. Researchers have proposed deep-learning-based skin cancer detection
systems based on these architectures.

2.2.1 Alexnet
Krizhevsky proposed an AlexNet, a CNN network that consists of five convolutions and
three fully connected layers. AlexNet wastrained using 60 million parameters and solved
an overfitting problem using dropout [Link] achieved top-1 and top-5 error rates
of 37.5 and 17.0 on ImageNet LSVRC-2010

7
 Chapter 2 – Literature Review

2.2.2 Resnet
AlexNet won an ImageNet 2012 competition consisting of eight layers. In deep learning,
more layers are added to improve performance and minimize the error [Link] more
layers results in a vanishing gradient in which the gradient becomes zeroand exploding
gradient in which the gradient becomes too large. He solved theproblem of exploding
and vanishing gradients by introducing a concept of skip [Link] skip connec-
tion bypasses some levels in between to link layer activations to subsequentlayers to
make residual blocks stacked together to create a ResNet architecture. The layercausing
a problem during training can be skipped avoiding exploding and vanishinggradients,
and helps train deep neural networks.

2.2.3 Densenet
Huang presented a densely connected convolution network in which each layer was con-
nected to every other layer. In DenseNet, all previous layers’ feature maps are utilized
as inputs for each layer, and that layer’s feature map isused as an input for all layers that
come after it. DenseNet architecture helped to solve thevanishing gradient problem and
allowed feature reuse. Connecting each layer with everyother also helped strengthen fea-
ture propagation and decrease the number of parameters required to train a deep neural
network.

2.2.4 Mobilenet
Howard presented a lightweight network, MobileNet, for mobile applications. Mo-
bileNet replaces the 3 × 3 convolution operation in standard CNN with the 3 × 3 depth-
wise convolution and 1 × 1 pointwise convolution operation; using depthwise separable
convolution operation instead of standard convolution operation helps in reducingthe
number of training parameters. The difference between standard convolution operatio-
nand depthwise separable convolution used in MobileNet

2.3 Related Works

Machine learning techniques have outperformed traditional dermatological methods in
image processing and lesion detection due to their ability to analyze vast amounts of
data and learn complex patterns. Neural networks, a key component of machine learn-

8
 Chapter 2 – Literature Review

ing, contribute significantly to this advancement. They are computational models de-
signed by by taking the neural networks in human brain as [Link] instance,
Viswanatha Reddy Allugunti et al. [3] developed a specific Convolution Nural Network
(CNN) model for skin lesion classification. Convolution neural networks are a type
of neural networks which are first trained by using large number of labelled images.
For this model, they have used the dataset from dermnetz for training and also for bet-
ter training, image augmentation techniques to create diverse training examples. This
model achieved an accuracy of 88.83. In 2020, Amirreza Mahboda et al. [9] developed a
Multi Stage Fusion CNN (MSM-CNN) model for skin lesion classification. Multi-stage
fusion CNNs integrate features from multiple stages of a convolutional neural network
to enhance performance. For this model, they have used the dataset from ISIC 2018
for training. For better training, image augmentation, bicubic interpolation are used to
create diverse training examples. This model achieved an accuracy of 96.3.

In 2019, Nudrat Nidaa et al. [11] developed a RCNN model which is a modification to
CNN model which considers only melanoma regions as ROI and remaining images as
background. For this model to work effectively the artifacts are removed and image is
enhanced, the ground truth images are binarized and region-labelling segmentation is
applied. This model localises the melanoma regions with regression confidence score.
This model is trained on the ISIC 2016 dataset, the regions with regression confidence
score greater 0.5 are considered for segmentation. This model achieved an accuracy of
94 with a specificity of 0.94. To make the previous RCNN model more accurate and effi-
cient, in 2020 Shunichi Jinnai et al. [7] developed a model called Faster-RCNN(FRCNN)
network which is a combination of a RCNN network with a Region Proposal Network
(RPN) is used which is designed for suggesting potential regions of interest in the lesion
and then Fast R-CNN takes those regions, extracts features, and classifies the objects
within those regions. Together, they form a more efficient and accurate object detec-
tion system compared to earlier methods like R-CNN. Also, VGG-16 (visual geometry
group-16) was used to extract features from input images which then are used as inputs
of RPV. This model got an accuracy of 91.5 and a PPV OF 85. In 2020, Jinen Daghrir et
al. [5] made a study by comparing 3 classifying methods. Firstly, preprocessing done by
image colour enhancement by splitting RGB colours and hair removal by Otsu method
and thresholding method and pixel classification is done by hierarchical clustering. For
lesion segmentation morphological snakes are used. Features of low level are extracted
from texture, descriptors and colour. For training process, a SVM classifier, KNN clas-
sifier and a CNN is proposed. For the CNN to train large dataset is required so Data

9
 Chapter 2 – Literature Review

augmentation is done on 512 lesions images from the ISIC dataset. The proposed CNN
has out-performed among the 3 classifiers and produced an accuracy of 85.5.

To modify this CNN architecture for better results, in 2020, Ni Zhenga et al. [14] de-
veloped an optimised whale optimisation algorithm (WOA) which can optimize the
weights(filters) by adjusting them during the training process, seeking values that en-
hance the network’s ability to recognize relevant features in the input data. Also, to re-
duce a drawback called ‘premature convergence’, Levy flight technique Is used, which
introduces a random walk with long steps, enhancing exploration in the solution space,
thereby promoting the algorithm’s ability to find diverse and potentially better solu-
tions. They achieved an accuracy of 0.87 with a sensitivity of 0.9 by testing this model
of datasets like dermis, dermquest. In 2021, Md Shahin Ali et al. [2] proposed a method
using transfer learning, where a pre-trained model on one task is used as a starting point
for a new task, leveraging the learned knowledge to improve performance with limited
data. For optimisation of the results data was pre-processed, art facts and noise are
removed, data reduction and normalization through Z-score normalization. Data aug-
mentation was done to increase the robustness of the model. This model was run on the
HAM1000 dataset. Instead of SoftMax layer sigmoid layer is used for Binary classifi-
cation and achieved a precision of 96.57 and recall of 93.66. []

There were many advancements to reduce the complexity of CNN and also increase its
efficiency at the same time. One such advancement is use of efficient net models. Ef-
ficient Net is a type of Convolutional Neural Network (CNN) that balances model size
and performance through a technique called compound scaling, achieving high accuracy
with fewer parameters. In 2022, Karar Ali et al. [6] investigated the classification perfor-
mance of the EfficientNets B0-B7 on the HAM10000 dataset of dermatoscopic images.
The dataset consists of 10015 images belonging to seven skin cancer classes. They have
used ImageNet pre-trained weights to perform transfer learning and fine-tune the CNNs
for the HAM10000 dataset. Precision, Recall, Accuracy, F1 Score, Specificity, Roc Auc
Score, and Confusion Matrices were calculated to evaluate the EfficientNets B0-B7 per-
formance on this imbalanced multiclass classification task. Along all models that were
evaluated, EfficientNet B4 proved to be the best, achieving an 87 percent F1 Score and
87.91 percent Accuracy. In 2023, Vipin Venugopal et al. [13] developed an efficient
V2M and efficient net B4 models for classification. They too used the method ‘Transfer
Learning’ for training. Pre-trained models like Efficient Net V2-M and Efficient Net
B3 are used, which have already learned a lot from big datasets like ISIC 2022. The

10
 Chapter 2 – Literature Review

extracted features are propagated into a neural classifier model consisting of two dense
layers of 500 and 256 neurons, respectively, followed by output layers for seven targeted
classes. By using these, they achieved a high accuracy of 97.62 along with an F1 score of
0.98 and AUC of 0.99 for efficient net V2M model. In 2021 A. Murugan et al. [10] pro-
posed a method incorporating various types of cancer classification. Preprocessing was
done by enhancing the image using median filter and image segmentation done based
on ROIs. Feature extraction is performed using Moment Invariants, GLCM, GLRLM.
Many classifiers were used for detection task include SVMs, Probabilistic Neural Net-
works, Random Forests and a combination of SVM and RF. This model used the ISIC
dataset and the SVM+RF combination has achieved the highest accuracy of 89.31 and
sensitivity of 88.56.

In 2020 Javeria Amina et al. [4] worked on a method, where the images were resized
into 240 × 240 × 3 and RGB colour code was converted into LAB colour space focus-
ing on luminance channel. Segmentation is done through Biorthogonal 2-D wavelet
transforms and the Otsu algorithm. To extract deep features indicative of skin lesion
characteristics, the pre-trained Alex Net and VGG-16 models are leveraged. The PCA
(Principal Component Analysis) reduces the feature vector’s dimensionality. Multiple
classifiers were employed in this study, and the classification process was evaluated us-
ing both 5-fold cross-validation and a 0.5 hold-out validation strategy. In 2022, Talha
Mahaboob Alam et al. [1] developed a method where they concentrated more on the en-
hancement of the dataset through data augmentation HAM1000 Data set was enhanced
and it was given to three distinct classifiers: AlexNet, Inception V3, and RegNet Y-320.
They used two different data sets namely balanced and imbalanced and the best results
were achieved by using the balanced data set with the classifier RegNet Y-320 with the
learning rate 0.001. They achieved Accuracy-91 F1-score-88.1 ROC-0.95. In 2023, Sk
Hasane Ahammad et al [12]. developed a skin classification model with resnet 50 ar-
chitecture. Firstly, for preprocessing, Scikit-image was used, which is an open-source
Python library for image processing tasks. Then the images were resized to dimensions
227 × 227 pixels to reduce processing time. Then they were given as input for resnet
50 network. ResNet-50 processes input images through convolutional layers, extract-
ing features using residual connections that bypass certain layers. These skip connec-
tions mitigate vanishing gradient issues during training. They have taken xiangya derm
dataset for the input and achieved an accuracy of 87. Masum Shah Junayed et al. [8]in
2021 have developed a deep CNN model for detection and classification of skin cancer.
They have collected a skin cancer dataset, size 800 images of four skin cancer classes

11
 Chapter 2 – Literature Review

that are, Actinic Keratosis, Basal Cell Carcinoma, Malignant Melanoma and Squamous
Cell Carcinoma. They have also used data augmentation techniques for increasing the
dataset at 5600 images that were split into training and test set for deep CNN models. A
deep CNN model has been proposed that has convolution layers for extracting features
and fully connected layers for classifying the cancer type. Regularization techniques
like batch-normalization and dropout helped to reduce the overfitting. It has achieved
an accuracy of 87.36 without augmentation and 95.98 with augmentation.

2.4 Problem Statements

The challenge in skin cancer detection stems from limited diagnostic accuracy and de-
layed interventions. Access barriers to specialized care necessitate automated, scalable
solutions. Ethical concerns, including privacy and bias mitigation, complicate deep
learning model development. The goal is to enhance accuracy, ensure global acces-
sibility, and seamlessly integrate models into clinical workflows for widespread im-
pact.

2.5 Proposed Solutions

This project proposes the development of a skin cancer detection system using deep
learning techniques. The ISIC dataset will be utilized, and after preprocessing, the
dataset will be divided into two classes for binary classification. An EfficientNet B0
neural network will be selected as the base model and will be fine-tuned on a specific
dataset. The trained model will then be tested on a separate dataset, aiming to achieve
a high level of accuracy in detecting skin cancer. This project aims to demonstrate the
potential effectiveness of deep learning models in medical image analysis, specifically
in the detection of skin cancer, which can potentially aid in early diagnosis and treat-
ment.

12
 Chapter 3
Proposed Approach
 Chapter 3 – Proposed Approach

3.1 Inrtoduction
In the field of medical imaging, transfer learning is gaining popularity due to its superior
performance [5, 36–38]. Utilizing pre-trained networks like EfficientNet B0 allows the
model to leverage its expertise and weights for tasks in the target domain, eliminating
the need for extensive training data. Instead, fine-tuning these pre-trained networks on
the target data yields exceptional results even with small datasets [5, 36–38]. In this
study, we opted for EfficientNet B0, a highly efficient convolutional neural network ar-
chitecture, known for its balanced depth, width, and resolution scaling. Developed by
Tan and Le, EfficientNet B0 achieves state-of-the-art performance while being com-
putationally efficient [40]. Its architecture, featuring depthwise separable convolutions
and efficient channel attention, makes it well-suited for medical image classification
tasks. Preprocessing, training, and testing with the EfficientNet B0 model, followed by
a thorough examination and comparison of the results, constitute the main stages of this
research.

3.2 Datasets
The proposed method is fine-tuned and trained with a large combined dataset from ISIC
2020 [34], ISIC 2019 [26,28,29], and ISIC 2018 [48] dermoscopic images. The data
with unique information across the dataset is captured during the data cleaning process.
The cleaning process filters the duplicated and invalid images across the dataset and
categorizes the data with the skin lesion. It is observed that the dermoscopic images
are labeled into seven major classes that are actinic keratosis (AK), basal cell carci-
noma (BCC), benign keratosis-like lesions (BKL), melanoma (MEL), melanocytic ne-
vus (NV), squamous cell carcinoma (SCC), and unknown [27]. The labels are provided
by the medical experts based on the nature of skin lesions. These seven classes can be
broadly categorized into malignant and benign lesions. The MEL, BCC, and SCC are
grouped as malignant class, and the AK

14
 Chapter 3 – Proposed Approach

Figure 3.1: Skin Lesion

3.3 EffecientNet B0

Figure 3.2: The architecture of EffecientNet b0 CNN

EfficientNet B0 consists of a stack of convolutional layers, including multiple blocks of

inverted residual blocks. The exact number of layers and stride sizes can vary depend-
ing on the specific implementation and the scaling coefficients used. However, typically,
EfficientNet B0 comprises around 19 layers in total, with the stride size varying across
different parts of the network. In general, the stride size is often set to 1 for most con-
volutional layers within the blocks to preserve spatial information, while certain layers
may have larger strides to downsample the feature maps. The stride size may also be
adjusted in later stages of the network to control the spatial dimensions of the feature

15
 Chapter 3 – Proposed Approach

maps. Overall, EfficientNet B0 achieves a balance between model depth, computational

efficiency, and receptive field size through its carefully designed architecture and scaling
coefficients.

3.4 Breif about methodology

In our proposed methodology, we initiated by acquiring the skin cancer dataset from
Kaggle, specifically the ISIC 2020 dataset. This dataset provides a diverse range of skin
lesion images, crucial for training a robust classification model. With the dataset in hand,
we meticulously preprocessed the images to ensure optimal quality and consistency.
This involved comprehensive steps such as noise removal to eliminate artifacts, image
enhancement techniques to improve clarity, and resizing to standardize dimensions for
efficient processing. Additionally, we employed data augmentation strategies to augment
our dataset, including techniques such as rotation, flipping, and scaling, to enhance its
diversity and robustness, crucial for effective model training.

Subsequently, we proceeded to train the EfficientNet B0 model, a state-of-the-art convo-

lutional neural network architecture renowned for its efficiency and performance. Our
binary classification task aimed to distinguish between melanoma and basal cell carci-
noma, two prevalent types of skin cancer with distinct visual characteristics. Leveraging
the Adam optimizer with a carefully chosen learning rate of 0.001, we embarked on the
training process over a total of 30 epochs. To monitor the model’s progress and prevent
overfitting, we implemented a validation frequency of 20 epochs, allowing us to assess
performance on a separate validation set throughout the training process.

Upon completion of the training phase, we rigorously evaluated the trained model’s per-
formance using a separate set of test images. By feeding the test images into the trained
network, we obtained predictions for each image and compared them against ground
truth labels to assess classification accuracy. We then meticulously analyzed and visu-
alized the results, examining metrics such as accuracy, precision, recall, and F1-score to
gain insights into the model’s strengths and weaknesses. These findings were compre-
hensively documented and presented in our project report, providing valuable insights
into the efficacy of the EfficientNet B0 model for skin cancer classification tasks.

16
 Chapter 4
Results and Discussion
 Chapter 4 – Results and Discussion

4.1 Introduction
At this point, training and testing are the two phases of analysis. In order to determine
which model is best, an analysis is conducted to evaluate and compare the models that
have been used. A comparison of speed ,accuracy of the model in categorising images
into multiple preprocessed classes is the purpose of the assessment. Only the accuracy
outcomes and the model’s training time are used to compare models during the pro
cess of training. This evaluation determines which model can classify images with the
highest accuracy. Additionally, a confusion matrix is used during the testing process to
more precisely assess and contrast the models’ [Link], precision, sensi
tivity and f1-score are the four performance measurements.—make up the assessment
parameters in testing

4.2 Evaluation and Discussion

Wehave evaluated the model’s output using multiple performance metrics. A machine
learning algorithm’s performance can be evaluated using performance metrics, which
are essential measurements [35]. Performance metrics are a useful tool for evaluating
multiple models’ performances on a given data set and determining which model pro
duces the best results. Metrics of performance may be utilised to assess the overall ac
curacy of the model and identify areas for [Link] it comes to deep learning
and transfer learning strategies, the most popular performance metrics are

1. accuracy

2. precision

3. sensitivity

4. f1-score

In the given dataset, we structured the data into two distinct classes: melanoma and basal
cell carcinoma. This categorization served as the foundation for training our model,
enabling us to develop a robust skin cancer detection system.

Through our methodological approach, we aimed to discern patterns and features unique
to each class, thereby enhancing the model’s ability to accurately classify skin lesions.
By training on this differentiated dataset, we sought to optimize the model’s sensitivity

18
 Chapter 4 – Results and Discussion

and specificity in distinguishing between melanoma and basal cell carcinoma, crucial
for precise diagnosis and treatment decisions.

The training data results of the trained EffecientNet BO are displayed in the graphs be-
low

Figure 4.1: EffecientNet B0 training progress

Both the models achieved 99.94% accuracy respectively,indicates that both models have
learned the training data very well. These high accuracies suggest that the models have
effectively captured the patterns and relationships present in the training dataset, lead-
ing to accurate predictions on the training examples. Following the training analysis,
measurements and evaluations were also conducted on the testing that is shown in Ta-
ble I There are four parameters used, namely accuracy, sensitivity , precision and f-
score

19
 Chapter 4 – Results and Discussion

Figure 4.2: Confusion matrix for the test results.

Upon reviewing the provided confusion matrix, we can glean insights into the perfor-
mance metrics of the classification model. By analyzing the True Positives (TP), False
Positives (FP), True Negatives (TN), and False Negatives (FN), we can derive key eval-
uation measures including accuracy, precision, recall, and F1 score.

20
 Chapter 5
Conclusion and Future Work
 Chapter 5 – Conclusion and Future Work

5.1 Conclusion and Future work

In the progression of the skin cancer detection project leveraging deep learning tech-
niques, significant strides have been accomplished up to the stage of image enhance-
ment and noise reduction. These preprocessing steps lay the foundation for subsequent
advanced processes. Moving forward, the crucial tasks of image segmentation, fea-
ture extraction, image classification, and cancer detection through Deep Neural Network
(DNN) techniques remain pivotal to achieving the project’s objectives.

The completion of image segmentation will enable the identification of regions of in-
terest within dermatoscopic images, providing a segmented view for subsequent anal-
ysis. Following segmentation, feature extraction will unveil essential characteristics
and patterns inherent in skin lesions, forming the basis for accurate classification. The
implementation of a DNN for image classification holds promise for its ability to au-
tonomously learn intricate representations from extracted features, thereby distinguish-
ing between benign and malignant lesions.

Moving forward, future work in the skin cancer detection project entails refining image
segmentation techniques for more precise lesion delineation and exploring advanced
methodologies for feature extraction to capture nuanced information crucial for clas-
sification. The optimization of Deep Neural Network (DNN) models is imperative, in-
volving fine-tuning parameters, exploring alternative architectures, and addressing inter-
pretability to enhance transparency in decision-making. Additionally, continual dataset
expansion to encompass diverse skin conditions and demographic groups is essential
for robust model performance. Rigorous validation on independent datasets, seamless
integration into clinical workflows, and ongoing attention to ethical considerations, in-
cluding patient privacy and bias mitigation, will ensure the responsible and effective
deployment of the developed deep learning model in real-world dermatological diag-
nostics.

22
 References

[1] Talha Mahboob Alam, Kamran Shaukat, Waseem Ahmad Khan, Ibrahim A
Hameed, Latifah Abd Almuqren, Muhammad Ahsan Raza, Memoona Aslam, and
Suhuai Luo. An efficient deep learning-based skin cancer classifier for an imbal-
anced dataset. Diagnostics, 12(9):2115, 2022.

[2] Karar Ali, Zaffar Ahmed Shaikh, Abdullah Ayub Khan, and Asif Ali Laghari. Mul-
ticlass skin cancer classification using efficientnets–a first step towards preventing
skin cancer. Neuroscience Informatics, 2(4):100034, 2022.

[3] Viswanatha Reddy Allugunti. A machine learning model for skin disease classi-
fication using convolution neural network. International Journal of Computing,
Programming and Database Management, 3(1):141–147, 2022.

[4] Javeria Amin, Abida Sharif, Nadia Gul, Muhammad Almas Anjum, Muham-
mad Wasif Nisar, Faisal Azam, and Syed Ahmad Chan Bukhari. Integrated design
of deep features fusion for localization and classification of skin cancer. Pattern
Recognition Letters, 131:63–70, 2020.

[5] Jinen Daghrir, Lotfi Tlig, Moez Bouchouicha, and Mounir Sayadi. Melanoma skin
cancer detection using deep learning and classical machine learning techniques: A
hybrid approach. In 2020 5th international conference on advanced technologies
for signal and image processing (ATSIP), pages 1–5. IEEE, 2020.

[6] Ulzii-Orshikh Dorj, Keun-Kwang Lee, Jae-Young Choi, and Malrey Lee. The skin
cancer classification using deep convolutional neural network. Multimedia Tools
and Applications, 77:9909–9924, 2018.

[7] Shunichi Jinnai, Naoya Yamazaki, Yuichiro Hirano, Yohei Sugawara, Yuichiro
Ohe, and Ryuji Hamamoto. The development of a skin cancer classification system
for pigmented skin lesions using deep learning. Biomolecules, 10(8):1123, 2020.

23
 [8] Masum Shah Junayed, Nipa Anjum, Abu Noman, and Baharul Islam. A deep cnn
model for skin cancer detection and classification. 2021.

[9] Amirreza Mahbod, Gerald Schaefer, Chunliang Wang, Georg Dorffner, Rupert
Ecker, and Isabella Ellinger. Transfer learning using a multi-scale and multi-
network ensemble for skin lesion classification. Computer methods and programs
in biomedicine, 193:105475, 2020.

[10] A Murugan, S Anu H Nair, A Angelin Peace Preethi, and KP Sanal Kumar. Di-
agnosis of skin cancer using machine learning techniques. Microprocessors and
Microsystems, 81:103727, 2021.

[11] Nudrat Nida, Aun Irtaza, Ali Javed, Muhammad Haroon Yousaf, and Muham-
mad Tariq Mahmood. Melanoma lesion detection and segmentation using deep
region based convolutional neural network and fuzzy c-means clustering. Interna-
tional journal of medical informatics, 124:37–48, 2019.

[12] Teck Yan Tan, Li Zhang, and Chee Peng Lim. Intelligent skin cancer diagnosis
using improved particle swarm optimization and deep learning models. Applied
Soft Computing, 84:105725, 2019.

[13] Vipin Venugopal, Navin Infant Raj, Malaya Kumar Nath, and Norton Stephen. A
deep neural network using modified efficientnet for skin cancer detection in der-
moscopic images. Decision Analytics Journal, 8:100278, 2023.

[14] Ni Zhang, Yi-Xin Cai, Yong-Yong Wang, Yi-Tao Tian, Xiao-Li Wang, and Ben-
jamin Badami. Skin cancer diagnosis based on optimized convolutional neural
network. Artificial intelligence in medicine, 102:101756, 2020.

24