See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/383944913

SpectraDerm Vision: Deep Learning for Multiclass Skin Cancer Analysis

BACHELOR OF TECHNOLOGY In Electronics and Communication Engineering

Thesis · September 2024

DOI: 10.13140/RG.2.2.32502.82248

CITATIONS READS

0 27

1 author:

Shambhavi Sinha
Manipal Academy of Higher Education
4 PUBLICATIONS 1 CITATION

SEE PROFILE

All content following this page was uploaded by Shambhavi Sinha on 11 September 2024.

The user has requested enhancement of the downloaded file.

SpectraDerm Vision: Deep Learning for
Multiclass Skin Cancer Analysis

A Graduate Project Report submitted to Manipal Academy of Higher

Education in partial fulfilment of the requirement for the award of the
degree of

BACHELOR OF TECHNOLOGY
In
Electronics and Communication Engineering

Submitted by
Shambhavi Sinha
Reg. No: 200907122

Under the guidance of

Dr. Anu Shaju Areeckal
Assistant Professor
Department of Electronics and Communication Engineering

DEPARTMENT OF ELECTRONICS AND COMMUNICATION ENGINEERING

MANIPAL-576104, KARNATAKA, INDIA

JULY/AUGUST 2024
 DEPARTMENT OF ELECTRONICS AND COMMUNICATION ENGINEERING

Manipal
st
1 July, 2024

CERTIFICATE

This is to certify that the project titled SpectraDerm Vision: Deep Learning for
Multiclass Skin Cancer Analysis is a record of the bonafide work done by Shambhavi
Sinha (Reg. No. 200907122`) submitted in partial fulfilment of the requirements for the
award of the Degree of Bachelor of Technology (BTech) in ELECTRONICS AND
COMMUNICATION ENGINEERING of Manipal Institute of Technology, Manipal,
Karnataka, (A Constituent Unit of Manipal Academy of Higher Education), during the
academic year 2023 - 2024.

Dr Anu Shaju Areeckal Dr. Pallavi R Mane

Assistant Professor, ECE HOD, ECE
M.I.T, MANIPAL M.I.T, MANIPAL

ii
 ACKNOWLEDGMENTS

I would like to express my sincere gratitude to the individuals whose support and guidance
were instrumental in the successful completion of this project.

First and foremost, I am deeply indebted to Director Anil Rana for his unwavering support and
encouragement throughout the course of this project. His leadership and vision have been a
source of inspiration.

I extend my heartfelt thanks to Dr. Pallavi R Mane, Head of the Department, for her valuable
insights, continuous support, and for providing an environment conducive to learning and
research.

A special note of thanks to Dr. Anu Shaju Areeckal for her expert guidance, constructive
criticism, and constant encouragement. Her expertise and advice were crucial in navigating the
challenges faced during this project.

I would also like to acknowledge the invaluable assistance of Manan Bhatt (Reg. 200907188).
His support and collaboration were significant in the successful completion of this work.

Finally, I am grateful to all the faculty members whose assistance and cooperation were sought
during the project work. Their knowledge, suggestions, and support have been immensely
helpful.

iii
 ABSTRACT

Skin cancer is among the most prevalent cancers worldwide, significantly affecting a large
portion of the global population. There is an urgent need for an accurate, non-invasive method
to differentiate between benign and malignant tumors from dermoscopic images. This paper
introduces a novel lightweight deep learning model, SpectraDerm Vision, which enhances the
accuracy of skin cancer diagnosis. This model integrates advanced preprocessing and color
consistency techniques with a new Dual Attention mechanism, based on the modified
MobileNet-V2 architecture. The effectiveness of SpectraDerm Vision is demonstrated using
the public ISIC Challenge dataset, where it achieves impressive results: 98.85% accuracy,
99.28% precision, 98.21% recall, and 98.74% F1-score. SpectraDerm Vision shows
significantly superior performance compared to existing transfer learning models, making it
highly effective for the precise detection of skin cancer from dermoscopic images.
Additionally, we have extended the experiment to multiclass classification using the 2019 ISIC
dataset for 8 classes, achieving remarkable results with a Macro Precision of 94.68%, Macro
Recall of 91.94%, and Macro F1 Score of 92.16%. These outcomes further underscore the
robustness and versatility of SpectraDerm Vision in accurately classifying various types of skin
lesions, making it a powerful tool for dermatological diagnostics.

iv
LIST OF TABLES

Table No Table Title Page No

1.1 Timeline 2
3.1 Dataset Split 9
4.1 Result for Macro Average Metrics 27
4.2 Result for Micro Average Metrics 27
4.3 Result for Weighted Average Metrics 28
4.4 Comparative Analysis of related works 30-31

v
 LIST OF FIGURES

Figure No Figure Title Page No

Illustration of preprocessing techniques applied on various
3.1 classes of malignant skin cancer images along with benign 14-17
samples.

Architecture of proposed Multiscale-FocusNet, comprising of

3.2 19
Feature extraction (MB Block), Multi-scale feature learning
(Block) and Final prediction (Final Block)
Architecture of proposed Feature extraction block – MB Block,
3.3 comprising activation layers, pooling layers and dual headed 20
attention layer at the end.
Architecture of proposed Multi-scale Feature Learning block,
3.4 comprising various convolution layers with stabilising dropout 21
and batch normalisation layers.

Dual Attention Module: Spatial attention coupled with Channel

3.5 attention, comprising two fully connected layers (FC1 and FC2) 22
and Global Average Pooling (GAP)

Final Block, comprising of Global pooling average and dense

3.6 23
layers to condense the information for prediction

Example of how sigmoid processes block of input values and

3.7 24
Ideal waveform function of SoftMax

3.8 SoftMax activation implemented in SpectraDerm 24

4.1 Confusion Matrix 28

vi
 Contents
Page
No
Acknowledgement i
Abstract ii
List Of Figures iii
List Of Tables vi

Chapter 1 INTRODUCTION
1.1 Introduction 1
1.2 Introduction to Area of work 1
1.3 Brief Present Day Scenario with Regard to the Work Area 1
1.4 Motivation 1
1.5 Organization of Report 2
1.6 Target Specification 2
1.7 Project Work Schedule 2
1.8 Organisation of Report 2

Chapter 2 BACKGROUND THEORY and/or LITERATURE REVIEW

2.1 Introduction 4
2.2 Literature Review 4
2.3 Literature Survey 4-5
2.4 Summary 6
2.5 General Analysis 6-7
2.6 Theoretical Discussion 7-8
2.7 Conclusion 8

Chapter 3 METHODOLOGY
3.1 Introduction 9
Methodology:
3.2.1: Dataset Used
3.2 3.2.2: Dataset Splitting 9-24
3.2.3: Addressing Class Imbalance
3.2.4: Preprocessing and Algorithm Description

vii
 3.2.5: Proposed Model Architecture
3.3 Preliminary Result Analysis 24-25
3.4 Conclusion 25

Chapter 4 RESULT ANALYSIS

4.1 Introduction 26
Result Analysis:
4.2.1: Macro Average Performance Metrics
4.2 4.2.2: Micro Average Performance Metrics 26-29
4.2.3: Weighted Average Perforamance Metrics
4.2.4: Confusion Matrix
4.3 Tabulated Results 29
4.4 Significance of the Result obtained 29
4.5 Deviation from Expected Results 30-31
4.6 Conclusion 31

Chapter 5 CONCLUSION AND FUTURE SCOPE

5.1 Brief Summary of Work 32
5.2 Conclusion of Project 32-33
5.3 Future Scope of Work 33-34

Chapter 6 HEALTH, SAFETY, RISK AND ENVIRONMENT ASPECTS

6.1 Brief Summary 35
6.2 Health, safety, risk involved in the Project 35
6.3 Concluding Remarks 36

REFERENCES 37-38
PROJECT DETAILS 39

viii
 CHAPTER 1
INTRODUCTION

1.1. Introduction
This chapter outlines the scope and intentions of the project, detailing the methodology
and the objectives aimed at enhancing the precision of skin cancer classification using
advanced deep learning techniques. It sets the stage for a detailed discussion on the current
landscape of skin cancer diagnosis, the motivation behind this project, and the expected impact
of the research.

1.2. Introduction to the Area of Work

Skin cancer, a prevalent and potentially lethal malady, necessitates advancements in
diagnostic methods to improve early detection rates [1]. This project focuses on the application
of deep learning to dermatoscopic images to classify skin cancer into 14 distinct categories.
The use of such technology represents a significant shift from traditional diagnostic methods
that are often invasive and subjective.

1.3. Brief Present Day Scenario with Regard to the Work Area
Currently, the diagnosis of skin cancer is primarily conducted through histopathological
examination, requiring tissue samples and expert analysis. However, this method is not only
invasive but also prone to delays and variability in diagnosis due to the subjective nature of
visual assessments and the scarcity of specialized dermatopathologists.

1.4. Motivation to Do the Project Work

Shortcomings in Previous Work: While existing models like MobileNet and other
transfer learning approaches offer a starting point, they often fall short in terms of
accuracy, speed, and the ability to handle diverse and complex image data.
Brief Importance of the Work in the Present Context: With rising skin cancer rates
globally, there is a pressing need for more accessible, accurate, and non-invasive
diagnostic methods.
Uniqueness of the Methodology: The proposed model, SpectraDerm Vision,
introduces a novel Dual Attention mechanism within a modified MobileNet-V2
architecture, aimed at significantly enhancing diagnostic precision.
Significance of the Possible End Result: The successful implementation of this model
could lead to faster, more accurate skin cancer screenings, reducing the burden on
healthcare systems and improving patient outcomes.

1
1.5. Objective of the Work
Main Work Objective: To develop and validate a lightweight, efficient deep learning
model capable of high-accuracy classification of skin cancer from dermatoscopic
images.
Secondary Objective: To compare the performance of the proposed model against
existing benchmarks, demonstrating improvements in speed and accuracy.
1.6. Target Specifications
Importance of the End Result: The project aims to achieve over 98% accuracy,
precision, and recall, making the SpectraDerm Vision model a viable tool for clinical
and remote diagnostic settings.
1.7. Project Work Schedule
A detailed timeline is laid out for the project, from data collection through to final
testing, ensuring structured progress and timely completion.

Table 1.1 Timeline

o Assesing dataset best suited for the project
o Assesing previous work done on the similar projects
Feb 2024
o Assesing best pre-processing techniques which extentuates the
best feature representation

o Initial model characterization

Mar 2024 o Model development for binary classification to establish
baseline results

o Optimisation of the binary classification model

Apr 2024
o Fine tuning hyperparameters to achieve most optimal results

o Multi-class model characterization

May 2024 o Model development for multiclass classification based on
results obtained from previous tests

o Optimisation of the multi-class classification model

Jun 2024
o Fine tuning hyperparameters to achieve most optimal results

Jul 2024 o Documentation

2
1.8. Organization of the Project Report (Chapter Wise)
Introduction: Overview of the project, its significance, and structure.
Literature Review: Analysis of current technologies and previous work in the field.
Methodology: Detailed account of the techniques and processes used in the project.
Results and Discussion: Presentation and discussion of the findings from the project
testing.
Conclusion and Future Work: Summary of the project, its impact, and potential future
directions.

3
 CHAPTER 2
BACKGROUND THEORY

2.1. Introduction
This section explores the theoretical framework and prior studies that form the basis of
our project. It offers an extensive review of the literature, capturing the latest advancements
and theoretical discourse in dermatological image analysis, with a special emphasis on skin
cancer detection.

2.2. Literature Review:

Recent Developments in the Work Area: Early detection of skin cancer is crucial in
dermatology due to its aggressive nature and significant implications on patient outcomes.
A variety of methods have been explored to enhance early detection capabilities.

Brief Background Theory: In the literature, multiple machine learning methods have been
explored for skin cancer detection, indicating an ongoing evolution of technology in this
field. Each method presents a potential for further improvement.

2.3. Literature Survey:

Artificial Intelligence Techniques: Ozkan et al. [2] explored four distinct machine
learning models: Support Vector Machines (SVM), Decision Trees, Artificial Neural
Networks (ANN) and K-Nearest Neighbors (KNN),. Among these, the ANN model
achieved the highest performance on the PH2 dataset with an accuracy of 92.5% and an
F1-score of 90.45%.

Preprocessing Innovations: Rebouças Filho et al. [3] introduced a novel preprocessing

approach that includes the removal of dermal hair and normalization of skin tone to
counteract the glare from gel used during examinations. Their approach using a structural
covariance matrix for feature extraction and a Multi-Layer Perceptron classifier resulted in
accuracies of 94.5% and 89.93% on the ISIC 2016 and ISIC 2017 datasets, respectively.

Hybrid Models: Toğaçar et al. [4] integrated auto-encoders with MobileNetV2 and spiking
neural networks, leading to a classification accuracy of up to 93.54% when combining
datasets. Meanwhile, Hosny et al. [5] utilized AlexNet with extensive data augmentation
techniques to achieve an average accuracy of 98.61% on the PH2 dataset, with a precision
of 97.73%.

4
Optimization Techniques: Sayed et al. [6] combined convolutional neural networks with
the Bald Eagle Search optimization technique and used the ROS method to address data
imbalances in the ISIC dataset. This model achieved high metrics across accuracy,
specificity, sensitivity, and F1-score, highlighting the effectiveness of their novel method.

Comprehensive Reviews: Magdy et al. [7] provided a detailed review of various machine
learning and deep learning models for various skin cancer classification. They highlighted
the VGG-16 with a KNN-PDNN classifier as exceptionally effective, albeit noting the
model's substantial size as a limitation [8].

Enhanced Architectures for Multiclass Classification: Thanga Purni et al. [9] introduces
a unique architecture optimized specifically for multiclass classification of skin cancer. It
leverages an ensemble of deep learning techniques, significantly enhancing the robustness
and accuracy of predictions. The model's extensive use of data augmentation also improves
its generalization capabilities, showcasing superior performance in accuracy and precision
compared to other state-of-the-art models. However, the complexity of EOSA-Net could
require substantial computational resources, potentially limiting its deployment in
resource-constrained environments.

Comprehensive Analysis of CNN Architectures: Akter et al. [10] explore a range of deep
convolutional neural networks, including MobileNet, DenseNet, InceptionV3, ResNet-50,
VGG-16 and Xception, for multiclass skin cancer classification. This study highlights the
high accuracy rates, particularly with InceptionV3, which reaches up to 90%. It also
discusses the benefits and limitations of using stacking ensemble models, noting issues with
overfitting as training accuracy consistently exceeds validation accuracy.

Innovative Use of Vision Transformer Networks: Arshed et al. [11] investigate the
application of Vision Transformer (ViT) networks in combination with convolutional
neural network-based pre-trained models. This novel approach takes advantage of the
strengths of both architectures, enhancing performance in multiclass skin cancer
classification. The study provides competitive results in benchmarking against various
advanced methods. However, the high computational demand of Vision Transformers and
challenges in model interpretability remain as significant concerns.

This work sets the stage for 'SpectraDerm Vision,' a new Dual Attention model built on the
modified MobileNet-V2 architecture, promising further improvements in multi-class skin
cancer diagnosis through advanced preprocessing and colour consistency techniques.

5
2.4. Summary
The review underscores a pivotal trend towards integrating deep learning with
traditional and novel optimization techniques to enhance the detection and classification of
skin cancer. These developments reflect a significant potential for improving diagnostic
accuracy and patient management.

2.5. General Analysis

This analysis delves into critical factors influencing the development and deployment
of AI-based diagnostic models in dermatology:

Model Efficiency and Computational Load: As previously mentioned, the computational

requirements of models like VGG-16 (528 MB) limit their practical application in mobile
and real-time settings. Efforts to create more efficient models, such as EOSA-Net, aim to
balance performance with computational demands but also highlight the ongoing need for
models that can operate within the resource constraints typical of many clinical
environments.

Real-world Application and Validation: The transition from controlled test environments
to real-world clinical settings remains a significant challenge. Models that perform well on
datasets like ISIC may not necessarily translate directly to clinical success without thorough
real-world testing and validation to ensure they meet rigorous clinical standards and can
function effectively within existing medical workflows.

Scalability: The scalability of AI diagnostic tools is critical. Models must adapt to varying
levels of healthcare infrastructure, from small clinics to large hospitals, ensuring that all
patients benefit from advancements in AI diagnostics regardless of the setting.

Data Privacy and Security: Making sure the security of patient data is paramount along
with keeping the data private, especially as models become more integrated into healthcare
systems. Compliance by medical regulatory bodies is essential step in making deployment
of AI in medical sector more feasible.

Interoperability: Integrating AI models with existing healthcare IT systems, such as

EHRs, is crucial for seamless clinical workflows. This interoperability facilitates better data
exchange and coordination across healthcare services, enhancing the overall efficacy of
patient care.

6
Bias and Fairness: It is necessary to address importance of biases in AI models to ensure
that they cater to diverse group of patients. This involves training models on diverse
datasets to prevent biases that could adversely affect diagnostic accuracy for certain
demographic groups.

Continuous Learning and Adaptation: The ability of AI models to learn and adapt
continuously from new data is a promising aspect that can keep them relevant and improve
their diagnostic precision over time.

Regulatory Approval Processes: Navigating the regulatory landscape is crucial for

bringing AI diagnostic tools to market. Understanding and complying with these processes
ensures that new technologies are both safe and effective before they are used in clinical
settings.

2.6. Theoretical Discussions

The theoretical foundations of advancements in skin cancer detection incorporate a variety
of key areas, each critical for the development of highly effective diagnostic models:

2.6.1 Advanced Preprocessing Techniques:

The quality of input images significantly impacts the accuracy of classification algorithms.
Innovations in preprocessing techniques are vital to ensure that images are optimized for
analysis. Techniques such as adjusting for variable lighting conditions, enhancing image
quality, and reducing external noise are essential. These adjustments help to accentuate
critical features in the images, which is crucial for the effective training and performance
of models. For example, as seen in the work of Rebouças Filho et al., the removal of dermal
hair and normalization of skin tones were pivotal in enhancing model accuracy on
challenging datasets like ISIC 2016 and 2017.

2.6.2 Generalization Across Datasets:

The ability of models to generalize effectively across various datasets remains a persistent
challenge in machine learning. Variations in dataset characteristics, such as image
resolution, demographic diversity, and annotation quality, can all significantly influence
model performance. To address these issues, it is essential to enhance the robustness of
models. This involves not only training models with varied data inputs but also
continuously refining and adapting them to address and correct discrepancies arising from
dataset diversity. For instance, the EOSA-Net model, as introduced by Thanga Purni et al.,
demonstrates how optimizing network architecture can help achieve better generalization
by incorporating an ensemble of deep learning techniques, which enhances both robustness
and accuracy across diverse datasets.

7
 2.6.3 Scalability and Adaptability:
As AI-based diagnostic models evolve, the scalability and adaptability of these systems
become crucial. Models must be designed to scale effectively across different healthcare
environments and adapt to changes in technology and medical practice. This includes being
able to update algorithms as new data becomes available, ensuring that models remain
current and effective.

2.6.4 Ethical and Regulatory Considerations:

The development and deployment of AI in dermatology must also navigate ethical and
regulatory landscapes. This includes ensuring that models do not perpetuate biases and that
they adhere to ethical standards concerning patient data and privacy. Additionally,
navigating the regulatory approvals for medical AI tools involves demonstrating not only
efficacy and safety but also the transparency of algorithmic decisions, which is crucial for
gaining trust and acceptance in clinical settings.

2.7. Conclusions
The chapter concludes by reiterating the importance of continued research and
innovation in the field of dermatological imaging, highlighting how the proposed SpectraDerm
Vision model is poised to advance the capabilities of skin cancer diagnostics.

8
 CHAPTER 3
METHODOLOGY

3.1. Introduction
This chapter presents the comprehensive methodology employed for the project
focused on classifying various skin cancer using advanced machine learning techniques. It
elaborates on the dataset used, preprocessing steps, the innovative architecture of the proposed
model, preliminary results, and draws conclusions from these analyses.
3.2. Methodology
• Dataset Used:
The ISIC (International Skin Imaging Collaboration) 2019 [12,13] dataset is a
comprehensive collection of dermoscopic images designed to support the classification
of skin lesions into various diagnostic categories. This dataset is utilized for training
machine learning models to enhance the accuracy of detection of skin cancer and its
classification. The dataset includes images from the ISIC challenges of 2018 and 2017,
resulting in a substantial compilation of 25,331 images.
This dataset supports the classification of skin lesions into multiple diagnostic
categories, which include:
• Melanoma(MEL)
• Basal Cell Carcinoma(BCC)
• Squamous Cell Carcinoma(SCC)
• Actinic Keratosis(AK)
• Dermatofibroma(DF)
• Melanocytic Nevus(NV)
• Vascular Lesion(VASC)
• Benign Keratosis (BKL- including solar lentigo, seborrheic keratosis, and lichen
planus-like keratosis)

Table 3.1 Dataset split

Class Total Train Validation Test
AK 867 624 69 174
BCC 3323 2393 266 664
BKL 2624 1889 210 525
DF 239 172 19 48
MEL 4522 3256 362 904
NV 12875 9270 1030 2575
SCC 395 284 32 79
VASC 253 182 20 51

9
 • Data Split
The dataset is randomly sampled into train set, validation set and test set. The split of
train-test data is 8:2, and the train-validation split within the training set is 9:1. The table
above shows the distribution of images across these splits for each diagnostic category.

• Addressing Class Imbalance

To address class imbalance in the dataset, image augmentation techniques are employed.
These techniques increase the size of the training dataset by introducing controlled
variations to the images. By randomly rotating images within a defined range, applying
zooming, shifting both horizontally and vertically, and performing flips, diverse real-
world scenarios of skin cancer are simulated. This augmentation strategy contributes
significantly to the robustness and generalization potential of the model, enhancing the
reliability and adaptability of the research findings.

• Preprocessing
The presence of hair in skin images can interfere with the accurate diagnosis and
classification of skin conditions, such as melanoma. Removing hair from these images
is crucial for improving the performance of machine learning algorithms used in medical
image analysis. This report outlines a method for hair removal using image processing
techniques.

Algorithm Description:
The algorithm for hair removal consists of several key steps:
• Convert Image to Grayscale: Hair removal is performed on grayscale images
to simplify processing.
• Apply Gaussian Blur: Gaussian blur is used to smooth the image, reducing
noise and small details that can interfere with hair detection.
• Thresholding: Thresholding helps in differentiating hair from the skin. This step
converts the grayscale image into a binary image where the hair appears as white
regions.
• Morphological Operations: Morphological operations, such as closing and
opening, are applied to refine the detected hair regions.
• Inpainting: The detected hair regions are inpainted using the surrounding skin
pixels to remove the hair from the image.

Detailed Steps:
• Convert Image to Grayscale: The input color image is converted to a
grayscale image using the cv2.cvtColor function.

10
 • Apply Gaussian Blur: Gaussian blur is applied to the grayscale image to smooth
it.
• Thresholding: Otsu's thresholding method is used to create a binary image
where the hair appears white on a black background.
• Morphological Operations: Morphological closing is applied to fill small holes
in the hair regions.
• Inpainting: The hair regions in the binary image are inpainted in the original
grayscale image.

Mathematical Explanation
The mathematical foundation of the hair removal algorithm can be broken down into
several key components:
• Grayscale Conversion: Grayscale conversion is an important image processing
technique which is used to reduce a colour image to a single channel image,
where each pixel merely represents shades of gray rather than colour. This
simplification is achieved by eliminating the hue and saturation information
while retaining the luminance. This is particularly useful in applications where
colour does not carry essential information, such as texture analysis, edge
detection, and, in your case, skin cancer detection, where structural properties
might be more informative than colour. The conversion is done using the
formula:

Y = 0.299R + 0.587G + 0.114B

where Y refers to the luminance, and R to red, G to green, and B to blue.

• Gaussian Blur: Gaussian blur uses gaussian function to reduce noise and
unnecessary details in images. This function is used to create a smoothing kernel
that is then convolved with the original image. The kernel's size and standard
deviation (sigma) determine the extent of the blurring. By using Gaussian blur,
high-frequency components—essentially sharp edges and noise—are filtered
out, making the image smoother. This is particularly beneficial before
performing operations like edge detection, as it helps to reduce the potential for
detecting false edges due to noise. The Gaussian function is defined as:

G(x, y) = (1 / (2πσ²)) * exp(-(x² + y²) / (2σ²))

where G(x, y) is the Gaussian kernel, and σ is the standard deviation. The
convolution of the image with this kernel results in a blurred image.

11
 • Otsu's Thresholding: Otsu's Thresholding is an automatic thresholding
technique that is used to convert a grayscale image into a binary image. There
are two classes of pixels(background and foreground) which when passed
through the algorithm forms bi-modal histogram. This is used to compute the
optimum thresholding for the two classes. The threshold is determined such that
the intra-class variance (the variance within each class) is minimized. This
method is extremely effective when the histogram of the image has a distinct
separation between the pixel intensity values of the foreground and background.

• Morphological Operations: Morphological operations are an algorithm

defined to operate on group of image processing operations which are classified
based on shapes. These operations apply a structuring element to an input binary
image, producing an output binary image. The most basic morphological
operations are dilation and erosion:
• Dilation is a pixel additive process used to populate boundaries of object
in an image. It is typically used to bridge small breaks or gaps in figures.
• Erosion removes pixels on object boundaries, effectively shrinking the
size of objects. It is useful for removing small noise and detaching two
connected objects. These operations are often used in sequence (like
opening, which is erosion followed by dilation, and closing, which is
dilation followed by erosion) to enhance specific image features.

• The closing operation is defined as a dilation followed by an erosion:

A • B = (A ⊕ B) ⊖ B
where A is the input image, B is the structuring element, ⊕ denotes dilation,
and ⊖ denotes erosion. This operation helps in filling small holes in the binary
image.
• Inpainting: The lost and deteriorated parts of images can be reconstructed by
Inpainting. It is commonly used to restore damaged photographs or artworks, remove
objects from images, or repair areas where data is missing. The Telea inpainting
technique, one method of achieving this, uses fast marching methods (a numerical
technique for solving boundary value problems) to propagate pixel values into the
regions to be inpainted. It works by iteratively replacing the pixel values in the damaged
region with a combination of surrounding pixel values, taking into account the
geometric and photometric properties of the image for seamless blending.

Before proceeding further, please note that the images depicted below have undergone
a series of preprocessing techniques, including grayscale conversion, Gaussian blur, Otsu’s

12
thresholding, morphological operations, and inpainting as described above. These images
represent various classes of carcinoma and are provided to illustrate the step-by-step process
of how masks are created and applied to each image. This visualization helps demonstrate the
transformation from raw dermoscopic images to enhanced versions, highlighting how each
preprocessing step clarifies and emphasizes critical features necessary for the accurate
classification and analysis of skin cancer lesions. By showcasing these processed images, we
aim to underline the effectiveness of our methods in improving diagnostic accuracy and visual
clarity, essential for reliable medical analysis.

13
 (a)
14
 (b)
15
 (c)
16
(d)
17
Fig. 3.1(a)-(d) Illustration of preprocessing techniques applied on various classes of malignant skin cancer images along with
benign samples. First column shows the original unprocessed images, second column displays black resembling features, third
column represents pure dark pixels, fourth column represents continuous extension of similar density pixels, fifth column
represents the combined mask made from rest of the masks and fifth column represents the net filtered output. a) Melanoma
and Melanocytic Carcinoma, b) Basal Cell Carcinoma and Aktinic Keratosis, c) Benign Keratosis and Dermatofibroma, d)
Vascular Lesions and Squamous Cell Carcinoma.

3.2.5 Proposed Model Architecture

Our proposed model, SpectraDerm Vision, is inspired by the efficient and
robust MobileNet-V2 [14] architecture, which is renowned for its balance of accuracy
and computational efficiency, making it ideal for mobile and embedded vision
applications. SpectraDerm Vision is tailored specifically for the classification of skin
cancer from dermoscopic images and is structured into four distinct blocks, each
designed to enhance the model’s performance in critical areas of image processing and
classification.
The proposed model comprises of four blocks, namely feature extraction (MB
Block), Dual attention, multi-scale feature learning (Block) and final prediction (Final
Block), as shown in Fig. 2.
The Feature Extraction Block employs modified MobileNet-V2's inverted
residuals and linear bottlenecks, using depthwise separable convolutions to reduce
computational cost and model size while preserving high-quality feature extraction.
This is followed by a Dual Attention Mechanism that enhances model sensitivity to
relevant spatial and channel-wise features critical for identifying various skin
lesions. Multi-Scale Feature Learning integrates features across different scales to
capture comprehensive details essential for accurately classifying diverse lesion types.
Finally, the Final Prediction Block compiles all refined features to output a probability
distribution of potential skin cancer classifications through a combination of pooling
layers and a SoftMax classifier, ensuring precise diagnostic predictions.

18
Fig. 3.2. Architecture of proposed Multiscale-FocusNet, comprising of Feature extraction (MB Block), Multi-scale feature
learning (Block) and Final prediction (Final Block)

Our model, SpectraDerm Vision, incorporates several advanced components designed

to optimize the detection and classification of multi-class skin cancer:
3.2.5.1 Feature Extraction Block (Modified MobileNetV2)
• Pointwise Convolutions: This layer increases the dimensionality of the
channels, setting the stage for more complex transformations. Pointwise
convolutions essentially prepare the data for subsequent depthwise processing
by transforming the input data's depth, allowing for more detailed and nuanced
feature extraction in later stages.
• Depthwise Convolution: Specifically targets spatial feature extraction by
processing each input channel independently. This approach significantly
reduces the computational burden as it separates the convolutional process into
simpler, channel-wise operations, thus allowing for efficient feature processing
with reduced model complexity and faster execution times [15].

19
 • Batch Normalization and ReLU Activation: Applied sequentially after each
convolution, batch normalization normalizes the outputs, stabilizing the learning
process by reducing internal covariate shift. This is complemented by ReLU
(Rectified Linear Unit) activation, which introduces non-linearity to the model,
enabling it to learn and represent more complex patterns and distinctions in the
input data.

Fig. 3.3 Architecture of proposed Feature extraction block – MB Block, comprising activation layers, pooling
layers and dual headed attention layer at the end.

3.2.5.2 Multi-Scale Feature Learning

• Standard and Dilated Convolutions: Utilizes a combination of regular and
dilated convolutions to extract and integrate features across multiple scales. This
is particularly vital for skin cancer diagnosis, where lesion sizes can vary
dramatically. Standard convolutions capture fine details, while dilated
convolutions, which works by expanding the receptive field without dimishing

20
 the resolution or increasing the number of parameters. The kernel used is filled
with spaces to increases the field size (atrous effect).
• Atrous Convolution: Enhances the model's ability to incorporate wider
contextual information without losing detail in the image's resolution. By
adjusting the dilation rate, atrous convolutions can gather information from a
larger area, helping the model to better understand and interpret broader spatial
relationships within the image, which is essential for effective lesion analysis
[17].

Fig. 3.4 Architecture of proposed Multi-scale Feature Learning block, comprising various convolution layers with
stabilising dropout and batch normalisation layers.

3.2.5.3 Dual Attention Module

• Channel Attention: Focuses the model's computational resources on the
most informative channels within the feature map. By weighting the

21
 relevance of each channel based on its content, this module ensures that
features which are more predictive of skin cancer characteristics are
emphasized during the learning process, thereby enhancing the overall
discriminative capability of the model [16].
• Spatial Attention: Directs the model’s focus to crucial spatial regions within
the image. This mechanism refines the feature maps to concentrate more on
areas likely to contain diagnostic features of melanomas, thus improving the
model’s accuracy by reducing distractions from less relevant regions.

Fig. 3.5 Dual Attention Module: Spatial attention coupled with Channel attention, comprising two fully connected layers
(FC1 and FC2) and Global Average Pooling (GAP)

3.2.5.4 Final Prediction Block

• Global Average Pooling and Connected Layers: Simplifies the output by
condensing the entire feature map to a single vector through global average
pooling. This vector, representing the most critical features extracted across
the entire image, feeds into the fully connected layers, which then utilize this
condensed information for the final classification.
• Training with Categorical Cross-Entropy Loss: Utilizes categorical cross-
entropy as the loss function, which is tailored for multiclass classification
tasks like identifying various types of skin cancer. This loss function
quantifies the difference between the predicted probability distributions and

22
 the actual category labels across multiple classes. By doing so, it drives the
model to refine its predictions across all categories, enhancing the accuracy
and precision in classifying different types of skin lesions.

Fig. 3.6 Final Block, comprising of Global pooling average and dense layers to condense the information for prediction

3.2.5.5 SoftMax:
• The SoftMax function is a mathematical function commonly used in
classification models, particularly in the final layer of neural networks, to
transform raw output scores from the model into probabilities. It works by
exponentiating and then normalizing the output scores, ensuring that the sum
of the probabilities of all output classes equals one. This function is ideal for
multiclass classification problems as it provides a clear, probabilistic output
for each class, making it easier to determine which class the input most likely
belongs to base on the highest probability.

23
 a)

Fig. 3.7 a) Example of how sigmoid processes block of input values, b) Ideal waveform function of SoftMax

• In the SpectraDerm Vision model, the output from the final block, which
contains the refined and processed features, is fed into the SoftMax function
to obtain the final predictions. This step is crucial as it converts the logits
(raw classification scores) from the final block into probabilities that sum to
one. These probabilities represent the model’s confidence in each class of
skin cancer, allowing for a clear and decisive classification based on which
class has the highest probability. This final step is essential for ensuring that
the model's predictions are both interpretable and actionable, facilitating
accurate diagnostic decisions in clinical settings.

Fig. 3.8 SoftMax activation implemented in SpectraDerm

3.3. Preliminary Result Analysis

Initial testing has confirmed that our model, SpectraDerm Vision, combines
lightweight efficiency with robust performance. With a total model size of just 9.3MB and
comprised of 95,362 parameters, it is exceptionally well-suited for deployment in resource-

24
constrained environments such as mobile devices or remote clinical settings where computing
power is limited.

Initial Performance Metrics:

• Accuracy: The model achieves a high accuracy rate of 98.85%, indicating that it
correctly identifies the presence or absence of melanoma in nearly all cases.
• Precision: At 99.28%, the precision metric shows that almost every instance predicted
by the model as melanoma was indeed melanoma, underscoring the model's reliability
in diagnosis.
• Recall: The recall of 98.21% demonstrates the model's capability to identify a high
percentage of all actual melanoma cases present in the test data.
• F1-Score: An F1-score of 98.74% highlights the balance between precision and recall,
confirming the model's overall diagnostic efficacy.
• Binary Cross-Entropy Loss: A low loss score of 0.015 indicates that the model’s
predictions are very close to the actual binary labels, reflecting high confidence in its
outputs.
As we expanded the model to include eight different classes of skin cancer, a more complex
and nuanced evaluation was necessary to accurately measure its expanded capabilities.
Expanded Multiclass Performance Metrics:
• Weighted Precision: The precision adjusted for class imbalance stands at 94.68%. This
weighted metric takes into account the varying importance of each class based on its
representation in the dataset, ensuring that each class contributes proportionally to the
overall metric.
• Weighted Recall: Similarly, the weighted recall of 91.94% reflects the model’s
effectiveness across all classes, factoring in the relative sizes of each class to ensure
fairness and accuracy in measurement.
• Weighted F1 Score: The weighted F1 Score of 92.16% continues to provide a balanced
measure between precision and recall across multiple classes, confirming that the model
performs consistently well across a diverse range of skin cancer types.

3.4. Conclusions
The SpectraDerm Vision model represents a significant advancement in the non-
invasive and early detection of skin cancer. Its low computational demand paired with high
accuracy offers a promising solution for early detection campaigns, potentially improving
prognosis through timely and precise diagnosis. The model's performance, enhanced by
advanced architectural features such as dual-headed attention and dilated convolutions,
demonstrates a successful balance between model complexity and diagnostic efficacy. This
chapter underscores the potential for SpectraDerm Vision to revolutionize dermatological
screenings by providing a tool that is both accessible and effective in real-world settings.

25
 CHAPTER 4
RESULT ANALYSIS

4.1. Introduction

This chapter delves into the evaluation of the SpectraDerm Vision model,
specifically its performance on the SIIM-ISIC 2019 Challenge Dataset. We employ a
detailed set of evaluation metrics to thoroughly assess the model’s capability in diagnosing
various classes of skin cancer. The results are showcased through both graphical
representations and detailed tables, providing a clear visual and statistical analysis of the
model's effectiveness. This section will also explore the significance of the findings,
pinpoint any unexpected outcomes, and explore potential reasons for such deviations.
To comprehensively evaluate our model, we utilize a set of diverse metrics that
collectively gauge the diagnostic accuracy of the SpectraDerm Vision model in
distinguishing among different skin lesion types from dermoscopic images. These key
metrics include sensitivity, specificity, accuracy, and the F1-score, each offering insights
into different aspects of diagnostic performance.

4.2. Result Analysis

Tabular Presentation of Results:

Average Metrics for Multiclass Evaluation: In the context of evaluating our
model, SpectraDerm Vision, which deals with multiple skin cancer classes, it's crucial to
assess its performance comprehensively using various average metrics. These metrics help
in understanding the model's effectiveness across all classes and addressing issues like class
imbalance. Here are the three main types of averages used to calculate performance metrics:

4.2.1 Macro Average: The Macro Average calculates the average of the metric for
each class independently and does not take class size into account. This method
treats all classes equally, regardless of their frequency in the dataset. It provides
a measure of the model's effectiveness across all classes and can highlight if the
model is underperforming in any smaller classes.

26
 Table 4.1 Result table for Macro average metrics

Metrics Results
Macro Precision 94.68%
Macro Recall 91.94%
Macro F1 Score 92.16%

4.2.2 Micro Average: It computes the metrics globally by aggregating the

contributions of all classes. This method considers the class imbalance by
aggregating the totals of true positives, false positives, and false negatives
across all classes before the calculation, making it particularly relevant in
datasets where some classes are significantly overrepresented.

Table 4.2 Result table for Macro average metrics

Metrics Results
Micro Precision 93.77%
Micro Recall 90.94%
Micro F1 Score 92.06%

4.2.3 Weighted Average: The Weighted Average adjusts the metrics to account for
the number of instances (support) of each class, which is essential for
imbalanced datasets. This average multiplies each class metric by the class’s
support, sums these products, and divides by the total number of instances
across all classes. This method ensures that classes with more instances have a
proportionately greater impact on the overall metric.

27
 Table 4.3 Result table for Macro average metrics

Metrics Results
Weighted Precision 94.76%
Weighted Recall 92.58%
Weighted F1 Score 92.14%

4.2.4 Confusion Matrix: The confusion matrix below offers a detailed breakdown of
the SpectraDerm Vision model’s performance across different diagnostic categories.
This matrix is an essential tool for visualizing the model's ability to correctly classify
each type of skin lesion, as well as identifying which types are most frequently confused
with others.

Fig. 4.1 Confusion Matrix depicting the predicted instances of each class from the dataset.

28
 Each column of the matrix is used to represent the predicted probabilities of each instance
of the class. Similarly, the row is used to represent the instances which actual class. This
setup allows us to easily observe not only the true positive rates for each class but also the
specific types of errors, such as false positives and false negatives, that the model is making.
This information is critical for understanding the model’s strengths and weaknesses in
classifying different types of skin cancers and guiding further improvements.

4.3. Explanation for Tabulated Results

The updated performance metrics from the SpectraDerm Vision model highlight its
strong capabilities across multiple classes of skin cancer, demonstrating exceptional precision.
This precision is crucial in minimizing false positives—a common challenge in medical
imaging analysis—ensuring that non-malignant cases are not misclassified. This accuracy in
classification helps in avoiding unnecessary medical interventions, which can be stressful and
costly for patients.
The model also exhibits a high recall rate, crucial for a multiclass classification system
where detecting each type of skin cancer accurately is vital for effective treatment plans. High
recall ensures that the majority of malignant cases, regardless of the cancer type, are correctly
identified, thereby supporting early and accurate cancer detection.
Furthermore, the weighted averages used in evaluating the model, such as the Weighted
F1 Score of 92.16%, indicate a strong balance between precision and recall across all cancer
classes. This balance is essential in clinical settings, where the implications of missing a
diagnosis (false negatives) or incorrect diagnosis (false positives) are significant. The model's
ability to maintain high performance metrics across various classes confirms its utility in
complex diagnostic scenarios, where differentiating between multiple types of skin lesions is
critical.

4.4. Significance of the Result Obtained

The performance of SpectraDerm Vision signifies a substantial advancement in skin cancer
diagnostics. Its high precision and recall are particularly crucial in the medical field, where the
accurate diagnosis can significantly impact patient treatment planning and outcomes. The
model's efficiency in processing and its lightweight architecture make it suitable for real-time
applications, potentially integrating into mobile devices used in telemedicine or remote
diagnostics.
The comparative analysis with existing models such as Alexnet, Resnet-50, Resnet-101, and
VGG-16 [18,19,20] highlights SpectraDerm Vision's superior performance. These
comparisons, detailed in Table 2, show that SpectraDerm Vision not only surpasses these
models in key metrics but also offers a more efficient solution due to its fewer parameters and
lower computational demands.

29
 4.5. Deviations from Expected Results & Justification:
While SpectraDerm Vision performs exceptionally well across most standard metrics,
such as sensitivity and F1-score, there is a noted slight deviation in achieving a lower loss value
than anticipated. This discrepancy could be attributed to the inherent challenges in
differentiating between certain benign and malignant lesions that exhibit similar dermatoscopic
features. These lesions can sometimes blur the distinct boundaries that the model relies on for
classification, leading to slight inaccuracies in the final loss calculations. Further refinement of
the feature extraction and classification layers could potentially reduce this loss, enhancing the
model’s ability to discern nuanced differences in skin lesions more effectively.

Table 4.4 Comparative analysis with related work

Feature
Specificity Sensitivity F1-Score Accura
Methods Extract Dataset Classifier
(%) (%) (%) cy (%)
or
Ozkan et al. [2] PH2 ANN 96.11 90.86 92.50
LBP PH2 MLM 98.80 98.33 98.34 98.33
Rebouças Filho et
GLCM PH2 MLM-NN 95.71 90.83 91.23 90.83
al. [3]
HU PH2 SVM 91.90 83.33 84.02 83.33
MobileNetV2
+ Autoencoder
ISIC 93.44 93.01 93.83 93.20
without using
Toğaçar et al. [4] SNN
MobileNetV2
ISIC + Autoencoder 95.15 95.36 95.68 95.27
using SNN
ISIC-
Sayed et al. [6] SqueezeNet 96.75 100 98.40 98.37
2020
PH2-
DCNN 83.33 72.92 80
Original
Hosny et al. [5] PH2 -
Augment DCNN 98.93 98.83 98.61
ed
KNN-
ISIC KNN- VGG16 99.75 99 99.37 99.38
PDNN
Magdy et al. [7]
KNN-
98.50 90.25 94.13 94.38
MobileNetV2
Thanga Purani et al.
ISIC EOSA-Net 93.05
[9]

30
 HAM10 InceptionNetV
90
000 3
HAM10
Akter et al. [10] Xception 88
000
HAM10
DenseNet 88
000
HAM10
Arshed et al. [11] ViT 92.16 92.14 92.17 92.14
000
ISIC- SpectraDerm
Proposed Method 91.21 92.74 94.85
2019 Vision

4.6. Conclusions:
The SpectraDerm Vision model has demonstrated remarkable effectiveness in the
classification of various types of skin cancer, consistently outperforming existing models
across key diagnostic metrics such as accuracy, sensitivity, and F1-score. The model’s novel
integration of dual-headed attention mechanisms and specialized convolutional blocks
significantly boosts its capability to accurately identify critical dermatoscopic features. Despite
its compact size, the model does not compromise on performance, establishing it as a
formidable tool for clinical application.
The architecture’s ability to maintain high precision while being lightweight makes it
especially suitable for deployment in resource-constrained environments, enhancing
accessibility and efficiency in skin cancer diagnostics. Looking forward, there is substantial
potential for further enhancements to the model. Improvements in training methods, dataset
diversity, and real-time processing capabilities could broaden its applicability, making
SpectraDerm Vision a pivotal component in advancing dermatological diagnostics.
This chapter not only underscores the current achievements of the SpectraDerm Vision model
but also sets the stage for further discussions on potential improvements and the broader
implications of deploying this technology in real-world clinical settings. The overarching goal
remains to enhance diagnostic accuracy and accessibility, thereby significantly improving
patient outcomes in the healthcare system.

31
 CHAPTER 5
CONCLUSION AND FUTURE SCOPE OF WORK

5.1. Brief Summary of the Work

This paper presented SpectraDerm Vision, an advanced, comprehensive diagnostic
pipeline for skin cancer, distinguished by its integration of a novel dual attention
mechanism. The core objective of this research was to substantially improve the precision
and reliability of skin cancer diagnosis by utilizing sophisticated image processing and
advanced machine learning techniques. The innovative approach was rooted in the
deployment of state-of-the-art preprocessing methods designed to optimize the visual
clarity and consistency of dermatoscopic images. Key techniques included the removal of
extraneous hair and the adjustment of color levels to ensure uniformity across all images,
which are crucial for the accurate analysis of skin lesions.
The technical foundation of SpectraDerm Vision was further reinforced by the
implementation of a variety of convolutional strategies—pointwise and depthwise
convolutions were used to efficiently and effectively extract and process spatial information
from the images. Atrous and dilated convolutions were also employed to expand the
receptive fields of the network layers, allowing the model to capture broader contextual
information without increasing the computational burden. This combination of
convolutions enabled the model to detect and delineate intricate image patterns that are
characteristic of different types of skin lesions, from benign moles to malignant melanomas.
These techniques were meticulously integrated into the SpectraDerm Vision model,
fostering a sophisticated hierarchical system of feature extraction and analysis. By
processing data through this advanced architectural framework, SpectraDerm Vision was
able to discern subtle differences between various skin conditions with a high degree of
accuracy. This capability is critical for the early detection and treatment of skin cancer, as
it significantly enhances the likelihood of diagnosing at a stage where the disease is most
treatable. Through the implementation of this refined diagnostic tool, the paper aimed to
set a new standard in the field of dermatological imaging, potentially revolutionizing the
approach to skin cancer screening and diagnosis.

5.2. Conclusions
SpectraDerm Vision has showcased remarkable capabilities in the multiclass
classification of various skin cancers, achieving a Weighted Precision of 94.68%, Weighted
Recall of 91.94%, and a Weighted F1 Score of 92.16%. These metrics are particularly
telling of the model’s proficiency in handling class imbalances and accurately
distinguishing between multiple types of skin lesions, a crucial feature for effective clinical
diagnostics.

32
 The application of a dual-headed attention mechanism significantly enhances the
model's ability to focus on relevant features within complex dermatoscopic images. This
refinement in focus aids in distinguishing between benign and malignant lesions across
different categories of skin cancer, enhancing diagnostic accuracy. The high weighted
scores across precision, recall, and F1 demonstrate that SpectraDerm Vision not only
performs well on average but also maintains consistent accuracy across various classes, an
essential factor in clinical settings where misclassification can have serious implications.
The efficacy of SpectraDerm Vision in a multiclass diagnostic context underscores
its substantial potential for integration into clinical decision-support systems, offering a
significant enhancement over existing state-of-the-art models. This performance boost is
particularly valuable in dermatological healthcare, where the accurate early detection of
skin cancer can dramatically improve patient management and outcomes.
Looking ahead, there is considerable scope for further enhancing SpectraDerm
Vision. Future work can explore expanding the model's capabilities to include even more
diverse types of skin conditions, further refining its classification algorithms to improve
specificity and sensitivity across underrepresented classes. Additionally, clinical trials in
real-world settings are necessary to validate the model's effectiveness in practical
applications and to gather feedback for iterative improvements. Ultimately, the goal is to
integrate SpectraDerm Vision into widespread clinical use, providing dermatologists with
a powerful tool for early and accurate detection of skin cancer, thereby enhancing the
quality of care and patient outcomes.

5.3. Future Scope of Work

5.3.1. Integration with Real-Time Imaging Devices: Future developments could focus
on integrating SpectraDerm Vision with real-time imaging devices used in
dermatological clinics. This integration would allow for immediate analysis and
diagnosis during patient examinations, potentially leading to faster and more
accurate treatment decisions. By processing images in real-time, the model can also
provide immediate feedback to clinicians, which could be used to retake images or
adjust examination techniques to improve diagnostic accuracy.

5.3.2 Incorporation of Additional Diagnostic Factors: While the model currently

excels in image-based diagnosis, integrating additional patient data such as genetic
factors, patient history, or demographic information could enhance its diagnostic
precision. Machine learning models that take a holistic view of patient health tend
to be more accurate and can provide more personalized treatment recommendations.

5.3.3 Adaptive Learning from New Data: Implementing continuous learning

algorithms could enable SpectraDerm Vision to adapt and improve over time

33
 without needing explicit retraining phases. By learning from new data as it becomes
available, the model can adjust to new trends in dermatological conditions,
variations in imaging technology, and evolving medical standards, maintaining its
relevance and accuracy.

5.3.4 Expansion to Other Dermatological Conditions: While the current focus is on

skin cancer, SpectraDerm Vision has the potential to diagnose other skin conditions.
Expanding the model to identify a broader range of dermatological diseases could
make it a more comprehensive tool for dermatologists, increasing its utility and
market value.

5.3.5 Enhanced Model Explainability: Increasing the explainability of the model can
build trust among clinicians and patients by making its decision-making process
more transparent. Techniques such as feature importance mapping and the
integration of explainable AI frameworks can help in illustrating why certain
diagnostic conclusions were reached, which is particularly important in medical
settings for justifying treatment decisions.

5.3.6 Collaboration with Biomedical Researchers and Clinicians: Collaborating

closely with biomedical researchers and clinicians can provide new insights that can
be used to refine the model. This collaboration can lead to a better understanding of
the clinical requirements and constraints, which, in turn, can drive the development
of tailored AI solutions that better meet the needs of end-users.

5.3.7 Development of a Mobile Application: Developing a mobile application for

SpectraDerm Vision could greatly increase its accessibility and convenience. Such
an application could be used in remote areas or for preliminary assessments in
primary care settings, potentially reducing the load on specialized healthcare
services and facilitating early skin cancer detection.

34
 CHAPTER 6
HEALTH, SAFETY, RISK AND ENVIRONMENT ASPECTS

6.1. Brief Summary

Throughout the development and deployment of SpectraDerm Vision,
meticulous attention was paid to health, safety, risk management, and environmental
considerations. This proactive approach ensured that the project not only adhered to
ethical standards but also minimized potential risks and environmental impacts
associated with the deployment of new medical technologies.

6.2. Health and Safety

6.2.1 Patient Safety: The primary concern in the development of SpectraDerm
Vision was patient safety. Rigorous testing protocols were established to
ensure that the model provides accurate diagnostics without adverse effects
on patients. This included extensive validation studies to confirm that the
model does not produce misleading results that could lead to inappropriate
treatments.
6.2.2 Data Security and Privacy: Ensuring the confidentiality and security of
patient data was paramount. The project complied with HIPAA and GDPR
regulations to protect patient data, using advanced encryption and secure
data storage solutions to prevent unauthorized access.
6.2.3 Technology Risks: The potential for technological failure was mitigated
through robust software engineering practices, including multiple phases
of testing and quality assurance checks. The model's performance was
continuously monitored to identify and rectify any issues promptly.
6.2.4 Operational Risks: Operational guidelines were developed to ensure that
healthcare providers could integrate and utilize SpectraDerm Vision
effectively within existing clinical workflows without disruption.
6.2.5 Resource Efficiency: The project prioritized resource efficiency,
particularly in the use of computing resources. SpectraDerm Vision was
designed to be lightweight and energy-efficient, minimizing its carbon
footprint and making it suitable for use in environments with limited
energy resources.
6.2.6 Sustainable Development: Considerations were also made to align the
project with sustainable development goals. This included using
environmentally friendly practices in the model's development and
deployment stages, and promoting responsible consumption of digital
resources.

35
6.3. Concluding Remarks
The health, safety, risk, and environmental aspects of SpectraDerm Vision
have been thoroughly managed to ensure that the technology not only advances
dermatological diagnostics but does so in a manner that is safe, secure, and
sustainable. As the technology moves forward, continuous improvement in these areas
will remain a priority to address any new challenges that arise and to adapt to
evolving standards and expectations in healthcare technology.

36
 REFERENCES

[1]. Skin Cancer (Including Melanoma) -Patient Version (2023) -

https://www.cancer.gov/types/skin, Last accessed on 14-03-2024.
[2]. Ozkan, I.A., Koklu, M.: Skin Lesion Classification using Machine Learning
Algorithms. Int J Intell Syst Appl Eng. 5(4) (2017) 285–289.
[3]. Rebouças Filho, P.P., Peixoto, S.A., Medeiros da Nóbrega, R.V., Hemanth, D.J.,
Medeiros, A.G., Sangaiah, A.K., de Albuquerque, V.H.C.: Automatic histologically-
closer classification of skin lesions. Computerized Medical Imaging and Graphics 68
(2018) 40-54. https://doi.org/10.1016/j.compmedimag.2018.05.004.
[4]. Toğaçar, M., Cömert, Z., Ergen, B.: Intelligent skin cancer detection applying
autoencoder, MobileNetV2 and spiking neural networks. Chaos, Solitons & Fractals
144 (2021) 110714. https://doi.org/10.1016/j.chaos.2021.110714.
[5]. Hosny, K. M., Kassem, M. A., Foaud, M. M.: Skin Cancer Classification using Deep
Learning and Transfer Learning. 9th Cairo International Biomedical Engineering
Conference (CIBEC), Cairo, Egypt. (2018) 90-93.
https://doi.org/10.1109/CIBEC.2018.8641762.
[6]. Sayed, G.I., Soliman, M.M., Hassanien, A.E.: A novel melanoma prediction model for
imbalanced data using optimized SqueezeNet by bald eagle search optimization.
Computers in Biology and Medicine 136 (2021) 104712.
https://doi.org/10.1016/j.compbiomed.2021.104712.
[7]. Magdy, A., Hussein, H., Abdel-Kader, R.F., Salam, K.A.E.: Performance Enhancement
of Skin Cancer Classification Using Computer Vision. IEEE Access 11 (2023) 72120-
72133. https://doi.org/10.1109/ACCESS.2023.3294974.
[8]. Keras Applications (2023)- https://keras.io/api/applications/, Last accessed on 14-03-
2024.
[9]. Thanga Purni, J.S., Vedhapriyavadhana, R., "EOSA-Net: A Deep Learning Framework
for Enhanced Multi-Class Skin Cancer Classification using Optimized Convolutional
Neural Networks," in Journal of King Saud University - Computer and Information
Sciences, 2024. https://doi.org/10.1016/j.jksuci.2024.102007.
[10]. M. S. Akter, H. Shahriar, S. Sneha and A. Cuzzocrea, "Multi-class Skin Cancer
Classification Architecture Based on Deep Convolutional Neural Network," 2022 IEEE
International Conference on Big Data (Big Data), Osaka, Japan, 2022, pp. 5404-5413,
doi: 10.1109/BigData55660.2022.10020302.
[11]. Arshed, M.A., Mumtaz, S., Ibrahim, M., Ahmed, S., Tahir, M. and Shafi, M.6, "Multi-
Class Skin Cancer Classification Using Vision Transformer Networks and
Convolutional Neural Network-Based Pre-Trained Models," in Information 2023, 14,
415. https://doi.org/10.3390/info14070415
[12]. International Skin Imaging Collaboration, SIIM-ISIC 2019 Challenge Dataset (2020).
https://doi.org/10.34970/2020-ds01.

37
[13]. Rotemberg, V., Kurtansky, N., Betz-Stablein, B., Caffery, L., Chousakos, E., Codella,
N., Combalia, M., Dusza, S., Guitera, P., Gutman, D., Halpern, A., Helba, B., Kittler,
H., Kose, K., Langer, S., Lioprys, K., Malvehy, J., Musthaq, S., Nanda, J., Reiter, O.,
Shih, G., Stratigos, A., Tschandl, P., Weber, J., Soyer, P.: A patient-centric dataset of
images and metadata for identifying melanomas using clinical context. Sci Data 8
(2021) 34. https://doi.org/10.1038/s41597-021-00815-z.
[14]. Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., Chen, L.C.: Mobilenetv2: Inverted
residuals and linear bottlenecks. In Proc. IEEE conference on Computer Vision and
Pattern Recognition (2018) 4510-4520. https://doi.org/10.1109/CVPR.2018.00474.
[15]. Chollet, F., Xception: Deep learning with depthwise separable convolutions. In Proc.
IEEE Conference on Computer Vision and Pattern Recognition (2017) 1251-1258.
[16]. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.,
Polosukhin, I.: Attention is all you need. Advances in Neural Information Processing
Systems (2017) 30.
[17]. Yu, F., Koltun, V.: Multi-scale context aggregation by dilated convolutions. In
International Conference on Learning Representations (ICLR) (2016).
[18]. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep
convolutional neural networks. Advances in neural information processing systems
(2012) 25.
[19]. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In
Proceedings of the IEEE conference on computer vision and pattern recognition (2016)
770-778.
[20]. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image
recognition. In ICLR (2015).

38
 PROJECT DETAILS

Student Details
Student Name Shambhavi Sinha
Register Number 200907122 Section / A/20
Roll No
Email Address [email protected] Phone No 7667388068
(M)
Student Name Manan Bhatt
Register Number 200907188 Section / B/27
Roll No
Email Address [email protected] Phone No 9769325418
(M)
Project Details
Project Title SpectraDerm Vision: Deep Learning for Multiclass Skin Cancer
Analysis
Project Duration 5 months Date of 1st July, 2024
reporting
Organization Details
Organization Manipal Institute of Technology
Name
Full postal Manipal Institute of Technology, Eshwar Nagar, Manipal, Karnataka,
address with pin 576104
code
Website address www.manipal.edu

Internal Guide Details

Faculty Name Dr. Anu Shaju Areeckal
Designation Assistant Professor
Full contact Dept of E & C Engg, Manipal Institute of Technology,
address with pin Manipal – 576 104 (Karnataka State), INDIA
code
Email address [email protected]

39

View publication stats