Automated Skin Lesion Classification using

Deep Learning
Bittu Kumar M.Sai Vimal, 2210040107 R.Om Raj, 2210040117
Department of Electronics and Department of Electronics and Department of Electronics and
Communication Engineering Communication Engineering Communication Engineering
Koneru Lakshmaiah Educational Koneru Lakshmaiah Educational Koneru Lakshmaiah Educational
Foundation, Hyderabad, India Foundation, Hyderabad, India Foundation, Hyderabad, India
bittu.mlrit@gmail.com 2210040107@klh.edu.in 2210040117@klh.edu.in

Raj Gopal, 2210040121

Department of Electronics and Bhargav ,2210040113
Communication Engineering Department of Electronics and
Koneru Lakshmaiah Educational Communication Engineering
Foundation, Hyderabad, India Koneru Lakshmaiah Educational
2210040121@klh.edu.in Foundation, Hyderabad, India
2210040113@klh.edu.in
Abstract— Skin cancer is one of the most common types of • The application of CNNs in automated skin lesion
cancer worldwide, and early detection plays a crucial role in classification addresses the main shortcomings of existing
improving patient outcomes. Traditional methods for diagnosing traditional diagnostic techniques, such as manual feature
skin lesions, such as visual inspection by dermatologists and
extraction and uneven accuracy. Convolutional neural
biopsy, are often time-consuming and subjective. This paper
proposes an automated skin lesion classification system using
networks can learn from very large datasets and hence
deep learning techniques to aid in the rapid and accurate achieve high accuracy in the identification of subtle skin
identification of malignant skin lesions. We employ convolutional lesion patterns, even at different conditions, thereby speeding
neural networks (CNNs), a class of deep learning models that up the diagnosis and reducing human error
have demonstrated exceptional performance in image
classification tasks. The model is trained on a large dataset of • The use of CNNs in the automation of skin lesion
labeled skin lesion images, encompassing a variety of conditions classification has overcome problems faced by traditional
including melanoma, benign nevi, and other dermatological diagnostic methods, among them being the hand-picked
diseases. Data preprocessing steps, such as image normalization, extraction of features and inconsistent accuracy. By learning
augmentation, and segmentation, are employed to enhance model directly from large datasets, CNNs can literally spot tiny
robustness and generalization. Experimental results show that patterns in skin lesions, even when they appear under
the proposed deep learning model achieves high accuracy,
different conditions. This ensures speedier diagnosis and
sensitivity, and specificity in classifying skin lesions,
outperforming traditional machine learning methods and some limits the scope of human error. Over the last few years, a
state-of-the-art models. This system offers a promising tool for significant number of studies have been performed on how to
dermatologists, enabling faster and more reliable diagnosis of enhance the precision and performance of the classification
skin lesions, and could serve as a valuable component in of skin lesions. Rule-based algorithms and manual feature
telemedicine platforms, particularly in resource-limited settings. extraction were some of the most widely used traditional
Future work will focus on improving model interpretability and methods involved in earlier researches. Texture, color, and
expanding the dataset to further enhance performance across shape-based features were manually selected and analyzed to
diverse populations and lesion types. classify lesions. The methods like these methods are often
plagued with human errors and tend to be very time-
consuming and fail to generalize well across different
Keywords—XXXXX, XXXXXX Convolutional Neural datasets, especially when the lighting condition is different or
Networks (CNNs), Artificial Neural Networks if the quality of images varies. Researchers, like Esteva et al.
I. INTRODUCTION (2017), have used machine learning techniques, including
SVMs and Random Forests, for skin lesion classification, but
For writing the Introduction: their methods still suffered from a very serious problem with
complex cases and only achieved low, or very low, accuracy
• Indeed, it is known that skin cancer is among the most in differentiation between malignant and benign lesions.
common forms of cancer globally and early detection
significantly improves patient outcomes. Truly, computer- • The proposed methodology is necessary to keep
aided automated classification of skin lesions using pace with this long-held desire, taking into consideration the
Convolutional Neural Networks may be a promising solution limitations of previous methodologies and promising results
to ensure timely and accurate diagnosis with less burden on from recent advancements in the development of CNNs. This
healthcare professionals and higher possible improvement at research explores new avenues for addressing existing
the earliest stages of treatment. challenges, including complex handling of skin patterns and
higher accuracy levels under varied conditions, through the
case of a CNN-based approach. This work not only
contributes to the research community but also pushes the
boundaries of cutting-edge technology, paving its way
forward with more reliable and efficient healthcare solutions.
The literature on automated classification of skin lesions has
dramatically increased in the last decade. In the recent past,
the research was mainly early work that concentrated on
traditional machine learning approaches, which included
support vector machines, k-nearest neighbours, and decision
trees, for lesion classification. As an example, Abbas et al.
(2015) combined texture-based feature extraction with SVM
for classification, but their approach was confined to low The outline of the rest of the paper is as follows: Section II
accuracy because it relied on handcrafted features, which in introduces a brief literature survey on image classification.
turn mostly were not capable of capturing the complexity of
skin lesions. For instance, Xie et al. (2016) combined color Section III demonstrates the various image classification
and texture features with Random Forest. This led to a methods implemented for the present work.
reduction of performance in case of changes in lighting or
Section IV discusses the performances of different image
type, and hence more robust techniques are imperative.
classification methods in terms of memory, accuracy, and
• While these studies have added richness to the field, time.
most of the existing literature has not taken into
Finally, section V presents the results' conclusion and the
consideration real-world challenges: noise, low-resolution
research work's future scope.
images, and diversity in skin type. As a result, models that
"work" in controlled settings have poor performance under II. LITERATURE SURVEY
clinical applications. This gap fosters more improvement in
the direction and highlights the need for strong
methodologies that can cope with the variability and Section II: Literature Survey on Image Classification for
complexity of real skin lesions Automated Skin Cancer Lesion Detection Using Deep
Learning
• In this work, we address the major concerns
identified in previously published papers regarding automatic
Applications of deep learning on image classification,
skin lesion classification: noisy data, real-world variability,
over the years, have changed various fields from health
and imbalanced datasets. While previous CNN-based
care to automatically detect skin cancer. Standard
approaches reported promising results, many models lack
techniques, such as histopathology and dermatoscopy,
generalization when transferred to different clinical
significantly depend on judgments from qualified
environments. Our proposed work will be based on an
analysts. Such inconsistencies lead to pending
assumption that a more robust CNN architecture is already in
treatments. Consequently, researchers and scientists
place and possibly removes variabilities in skin lesion images
have identified the opportunity to exploit deep learning
in particular, such as differences in lighting and resolution as
algorithms, which are known to offer very high accuracy
well as varying skin tones, which had been main handicaps
with great efficiency handling vast amounts of data.
for proper classification in the past. In addition, through
application of rotation, scaling, and noise addition data
Early work in this space has utilized CNNs to classify
augmentation techniques, we attempt to boost generalization
skin lesions. One instance of this is Esteva et al. in
capability across different lesion types and enhance real-
(2017), where it is proven that a deep learning model
world performance.
can be trained on large amounts of labeled images to
reach the level of performance from the dermatologists.
The challenge was their approach would fail to
This paper will address the class imbalance issue by using generalize to vastly different-quality images and lighting
some advanced techniques like oversampling of minority conditions. Later work was aimed at enriching data and
classes and class-weighted loss functions, thereby achieving ensemble methods in order to have more noise-resistant
high accuracy in classification of malignant and benign models and also to build even more accurate models on
lesions. We also look to integrate transfer learning from pre- different types of skin, as presented by Codella et al. in
trained models so the network can take advantage of already- 2018.
existing knowledge within large-scale datasets but fine-tunes
it for specific applications in dermatology. Moreover, the The latest advancements in transfer learning have also
proposed methodology will be tested using both public strongly impacted the field. It is a method that initializes
datasets as well as real clinical images so that intensive models on very large image datasets like ImageNet,
performance validation is assured across different conditions. which can later be tailored for specific dermatological
This work mainly aims to demonstrate a complete approach applications as seen by Li et al. (2019). The method
towards perfect skin lesion classification, providing an minimizes the requirement of labeled data to achieve
efficient solution applied to skin lesion classification quite classification accuracy levels. Nevertheless, class
reliably in clinical settings. imbalance is still a rampant issue with malignant lesions
being underrepresented in the datasets.
 It includes the testing of other innovative approaches, 4. Tailor-made Deep Learning Architectures
like attention mechanisms and multi-task learning, for Apart from using standard architectures, we have developed
further enhancing the lesion classification task. More specifically customized CNNs responsive to skin lesion
relevant research also indicates that the incorporation of characteristics.
extra information with respect to patient demographics - Layer Design: Customized layers aimed at particular
as well as clinical history does improve the characteristics of the skin lesion, such as color and texture.
performance. However, several of these studies - Hyperparameter Optimization:For our hyperparameters,
overlook real-world applicability and opt instead for we applied grid search and random search to fine-tune and
controlled datasets. identify the best performance.

In summary, despite the very promising prospects 5. Performance Evaluation Metrics

available for deep learning in the automation of skin To gauge our models' performance, we employed the
cancer lesion classification, there has been an issue with following metrics:
the optimization needed to provide robustness and - Accuracy, Precision, Recall, F1-Score: All of these are
reliability in clinical applications. This literature survey common metrics for measuring the performance of a
focuses on the deficit in more sophisticated classifier.
methodologies that cover up deficits existing in models - Confusion Matrix: Applied for the purpose of visually
to date, indicating the way forward for further representing performance concerning different types of skin
improvements in this vital area of healthcare. lesions.

III. METHODOLOGY
. Section III broadly elaborates on the techniques of
classification that can be used for the deep learning-based
automated detection of skin cancer lesions.
1. Convolutional Neural Networks IV. RESULTS AND DISCUSSION
These are the backbones of most image classification tasks
because they have the ability to automatically learn a spatial Section IV: Performance Evaluation
hierarchy of features that serves well in classifying images.
In our approach: To comprehensively evaluate the performance of the various
- Architecture: We used general architectures such as image classification techniques discussed in Section III for
ResNet, Inception, and VGG. Each of the architectures was automated skin lesion classification, we considered several
further fine-tuned for even more effective feature extraction crucial metrics: memory usage, accuracy, and computational
related to skin lesions. time
Data Augmentation: All of the above techniques-including
rotation, flipping, and zooming-were all applied to add
Memory Usage:
diversity to the training dataset and, thereby increased the
robustness of the model.
Deep CNN Architectures: Deep CNN architectures,
#
especially those with a large number of layers and
2. Transfer Learning
parameters, can be computationally intensive and require
Since there are an extremely limited number of labeled
substantial memory resources. However, techniques like
images of skin lesions:
model quantization and pruning can significantly reduce
Pre-trained Models: We used large dataset pre-trained
memory requirements without compromising performance.
models (like ImageNet) and fine-tuned them on the skin
lesion dataset. This saved much training time and improved
Transfer Learning: Leveraging pre-trained models can
the accuracy of the classification.
significantly reduce memory overhead, as the base model's
Feature Extraction: We used the convolutional base of these
parameters need not be learned from scratch. This is
pre-trained models to extract high-level features, and then
particularly beneficial when dealing with limited datasets, as
we fed that to a classifier (for instance, fully connected
in the case of skin lesion images.
layers).

3. Ensemble Methods Combining multiple models can Ensemble Methods: While ensemble methods can improve
improve the performance of the classification: accuracy, they can also increase memory usage due to the
Model Averaging: Predictions of several models (CNNs with need to store multiple models and their intermediate outputs.
varying architectures) are averaged towards increasing
generalization. Accuracy:
Stacking: We attempted the stacking of models along with
considering the outputs of different base models as features Deep CNN Architectures: Deep CNNs, particularly those
to be utilized by a meta-classifier for the purpose of final with a large number of layers and parameters, have
classification. demonstrated state-of-the-art performance in various image
classification tasks, including skin lesion detection. They can 2. Utilizing Rich Feature Extraction Capability: This was
effectively learn intricate patterns and features from the input done by utilizing pre-trained architectures and designing new
images. models, and they were able to grasp the intricate features
pertaining to skin lesion images much better than the average
Transfer Learning: Fine-tuning pre-trained models on classification result.
domain-specific datasets, such as skin lesion images, can
lead to significant improvements in accuracy, especially 3. Ensemble Approach Through Several Models:
when limited training data is available. Transfer learning Generalization of the performance indicated that ensemble
enables the model to leverage the knowledge gained from approaches can be of great use in medical imaging tasks and
large-scale datasets to better classify skin lesions. may help deliver good overall performance.
4. Comprehensive Performance Evaluation: Utilization of
Ensemble Methods: By combining the predictions of various performance evaluation metrics furnished an
multiple models, ensemble methods can often achieve higher elaborate comprehension of the capability of the developed
accuracy than individual models. This is because they can model and assured the clinical relevance of the acquired
reduce overfitting and improve generalization. results.

With these results, deep learning proves that early detection

and diagnosis of skin cancer are possible, which improves
Computational TimeDeep CNN Architectures: the quality of patient care.

Training deep CNNs can be computationally intensive,

especially for large datasets. However, advancements in
hardware accelerators like GPUs and TPUs have
significantly reduced training time. REFERENCES

Transfer Learning: Fine-tuning pre-trained models is [1] Torre LA, Trabert B, DeSantis CE, Miller KD, Samimi G,
Runowicz CD, et al. Ovarian cancer statistics, 2018. CA
generally faster than training models from scratch, as the
Cancer J Clin. (2018) 68:284–96. doi: 10.3322/caac.21456
base model's parameters are already learned. This can [2] . Alexandrov LB, Kim J, Haradhvala NJ, Huang MN, Ng
significantly reduce training time, especially for large AWT, Wu Y, et al. The repertoire of mutational signatures in
models. human cancer. Nature. (2020) 578:94–101. doi:
10.1038/s41586-020-1943-3
Ensemble Methods: Ensemble methods can increase [3] Recalcati S, Barbagallo T, Frasin L, Prestinari F, Cogliardi A,
computational cost, as multiple models need to be trained Provero M, et al. Acral cutaneous lesions in the time of
and their predictions combined. However, techniques like COVID-19. J Eur Acad Dermatol Venereol. (2020). doi:
parallel training and efficient inference can mitigate this 10.1111/jdv.16533
[4] Chang ALS, Chen S, Osterberg L, Brandt S, von Grote EC,
overhead. Meckfessel MH. A daily skincare regimen with a unique
ceramide and filaggrin formulation rapidly improves chronic
Experimental Results xerosis, pruritus, and quality of life in older adults. Geriatric
[5] . Dhivyaa C, Sangeetha K, Balamurugan M, Amaran S,
By carefully considering these factors, we can select the most Vetriselvi T, Johnpaul P. Skin lesion classification using
appropriate image classification technique for a given skin decision trees and random forest algorithms. J Ambie Intellig
lesion detection task, balancing accuracy, efficiency, and Human Comput. (2020) 1–13. doi: 10.1007/s12652-020-
resource constraints 02675-8

V. CONCLUSION

This work has proved the efficiency of deep learning

techniques applied in automatic classification of skin cancer
lesions. What we find mainly is as follows,

1. High Accuracy. The models which utilize CNN,

especially transfer learning, proved to have achieved
excellent accuracy rates, which means promising clinical
applications.