Melanoma Cancer Detection Using Deep Learning

G. Chinna Pullaiah, Vyshnavi Manchikanti, Shaguptha Naaz Dudekula,

Ravi Teja Mekalappagari and Viswa Teja Devarakonda

Computer Science and Engineering, Srinivasa Ramanujan Institute of Technology, Rotarypuram Village, B K Samudram

Mandal, Anantapur, Andhra Pradesh 515701, India

Keywords: Melanoma, Convolutional Neural Networks (CNNs), Dermoscopic Image Analysis, MobileNetV2, Deep

Learning.

Abstract: This study explores the skin behaviour and the fact that skin cancers, especially melanoma, can be fatal;

however, early detection can significantly improve the patient’s survival. This study presents a new approach,

which integrates image analysis with clinical information to improve the reliability of melanoma diagnosis.

Currently, dermatologists take dermoscopic photographic images of a skin lesion using a high-speed camera

and obtain a diagnostic accuracy of 65-80%. In case of additional specialist evaluations, this can increase to

75-95%. This paper uses CNNs, specifically the MobileNetV2, for skin disease subtype classification. It also

utilizes Linear Discriminant Analysis for their severity levels according to clinical data. The best performing

accuracy for the hybrid approach was achieved using CNN, with 92.32%, higher than that with traditional

image-only methodology. From being a simple custom- made application to user-friendly web application

using Flask is now been developed for real-time detection to avoid manual process and reduce time period for

detecting the type of melanoma. The fusion of AI technical platform and clinical curative, in this work

presented, provides a viable framework for early preliminary diagnosis of melanoma, thereby promoting

success and access to the healthcare system.

1 INTRODUCTION

The healthcare industry has seen exceptional

advancements in cancer detection in recent years,

much appreciated to modern innovations, developing

personal preferences, and rapid advancements in

artificial intelligence. As conventional person styles

meet ultramodern computerized advances, the field

faces the challenge of adjusting to a complicated

mix of restorative and innovative changes that are

reshaping how we diagnose, analyze, and treat

conditions (A. Esteva et al., 2017). To effectively

explore this changing geology, it’s significant to have

a profound understanding of the colorful variables

affecting this energetic territory, along with the

capability to fete and seize modern openings as they

emerge.

One of the major changes in cancer research is the

development and application of AI-based detection

systems. Conventional styles like visual checks and

biopsies, which were once in the past the standard, are

presently being outperformed by advanced education

methods that grant faster, more adaptable, and

increasingly exact choices. Developments such as

convolutional neural systems (CNNs), MobileNetV2,

and combined profound proficiency models have

revolutionized the field, driving to a modern period

where early and direct cancer detection is becoming

more widely accessible (T.J. Brinker et al., 2018).

Technological advancements play a crucial role in

determining how we diagnose medical conditions.

With high-quality dermoscopic pictures promptly

accessible and the utilize of cloud computing for

analyzing information, therapeutic pictures can

presently be reused more smoothly across diverse

stages. Modern innovations comparable to resolvable

AI and unified proficiency might assist improve how

specifically we analyze conditions and increment

croakers' belief in these frameworks by making AI

more straightforward and agreeable (D. Moturi et al.,

2024). Combining these innovations with restorative

information can create modern openings for early

disclosure of conditions and encourage

individualized treatment approaches.

As the field of diagnostics advances, we’re seeing

conventional healthcare styles alter altogether.

Pullaiah, G. C., Manchikanti, V., Dudekula, S. N., Mekalappagari, R. T. and Devarakonda, V. T.

Melanoma Cancer Detection Using Deep Learning.

DOI: 10.5220/0013908500004919

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 4, pages

81-87

ISBN: 978-989-758-777-1

Presently, Hospital-based diagnostics are confronting

competition from AI-driven telemedicine platforms,

which require a cautious approach to keep up both

delicacy and simple access for cases. Modern

computerized frameworks for identifying carcinoma,

often available through web operations, are pushing

healthcare providers to reevaluate their standard

clinical forms and incorporate AI tools into their work

(X. Lu et al., 2022). At the same time, experimenters

are looking into new applications for these advances,

like prognosticating persistent issues, covering how

well medications are working, and culminating

integration with electronic health records (EHRs).

The boundary between human clinicians'

expertise and AI-powered insight is becoming

increasingly blurred, highlighting the convergence of

medical knowledge and technology. Conventional

healthcare affiliations aren't well-adjusted to this

alteration; they're bouncing into the advanced

transformation, working difficult to keep up their role

in making judgments while also taking advantage of

what machine learning has to offer (L. Wei et al.,

2020). The around the world projection of AI personal

devices brings both instigative conceivable outcomes

and noteworthy challenges. These calculations need

to be planned to consider diverse skin types, designs

of complaint that change by locale, and the contrasts

in healthcare frameworks around the world (A. Ech-

Cherif et al., 2019). The most successful AI

implementations will be those that achieve high

accuracy across diverse populations while addressing

critical ethical concerns and adhering to necessary

regulations (M.Q. Khan et al., 2019).

Healthcare providers and researchers are

navigating a complex yet promising landscape, where

patient outcomes are of paramount importance (A.B.

Ali et al., 2016). There’s a parcel of plutocrat being

poured into idealizing how we clergyman datasets

and update calculations to keep up with the including

requirements for dependable AI diagnostics. Right

presently, there’s a" delicacy race" passing among

investigation teachers and tech companies to create

the a la mode carcinoma revelation models that can

work well in distinctive clinical environments. At the

same time, conventional person styles are being

improved through mutt models that mix the moxie of

croakers with AI perceptivity. This approach takes

advantage of times of restorative information,

whereas drinking modern developments (S. Bharathi

et al., 2021 and S. Bhadula et al., 2019).

This investigation examines the future of skin cancer

detection in the rapidly advancing field of AI-driven

diagnostics by assessing the innovative and clinical

variables affecting its relinquishment and adequacy.

The related works are listed in Section 2. The

suggested techniques are introduced in Section 3.

Section 4 reports the results. The discussion is given

in section 5. The last section contains the

conclusion.

2 RELATED WORKS

Study Novoa et al, 2017 presented a deep learning-

based melanoma detection system using

convolutional neural networks (CNNs) that achieved

89% accuracy in classifying dermoscopic images,

demonstrating the potential of AI in early skin cancer

diagnosis.

Research T.J. Brinker et al., 2018 examined a

transfer learning approach with MobileNetV2 for

skin lesion classification, showing improved

performance over traditional methods while requiring

less computational resources for medical image

analysis.

Author D. Moturi, et al, 2024 developed an

ensemble model combining CNN and SVM for

melanoma detection, achieving 91.3% accuracy on

the ISIC dataset and highlighting the importance of

multi-feature analysis.

Article X. Lu, et al, 2022 investigated a hybrid

deep learning system incorporating clinical metadata

with image data, resulting in a 7% improvement in

melanoma classification accuracy compared to

image-only models.

Paper L. Wei, et al, 2020 examined a federated

learning framework for melanoma detection that

preserved patient privacy while maintaining 88%

diagnostic accuracy across multiple healthcare

institutions.

Study A. Ech-Cherif, et al, 2019 proposed a vision

transformer (ViT) based approach for skin cancer

classification, demonstrating comparable

performance to CNNs while offering better

interpretability of decision- making processes.

Research M.Q. Khan et al., 2019 analyzed the

impact of different image augmentation techniques

on melanoma detection accuracy, finding that

geometric transformations combined with color

adjustments improved model robustness by 12%.

Article A.B. Ali, et al, 2016 created a lightweight

CNN architecture optimized for mobile deployment,

enabling real-time melanoma screening with 86%

accuracy on smartphone-captured images.

Paper S. Bharathi et al., 2021 investigated the use

of attention mechanisms in deep learning models for

melanoma detection, showing significant

improvement in identifying small and early-stage

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

lesions.

Paper Juyal et al., 2019 studied a multi-task learning

system that simultaneously performed lesion

segmentation and classification, achieving state-of-

the-art performance on both tasks for automated skin

cancer diagnosis.

Study explored an explainable AI framework for

melanoma detection that provided visual

explanations of model decisions, increasing clinician

trust in the system's predictions.

Research examined a 3D CNN approach for

analyzing sequential dermoscopic images of evolving

lesions, demonstrating improved accuracy in tracking

melanoma progression over time.

3 METHODOLOGY

3.1 Objective

This project aims to develop a deep learning-based

system for detecting melanoma, a type of skin cancer,

using advanced AI technology. This system will

analyze images of the skin to help diagnose cancer

more quickly and accurately. It utilizes advanced

deep learning techniques, particularly convolutional

neural networks (CNN), MobileNetV2, and hybrid

LSTM models, to ensure both precision and

efficiency in its calculations. Designed for use in

hospitals and remote healthcare settings, the system

can evaluate high-quality images of skin lesions. It

identifies key diagnostic features and provides

automated classifications, while also being able to

explain its reasoning in understandable terms. By

making the diagnosis process less subjective and

more accessible, this innovative system aims to

enhance detection rates, reduce false positives and

negatives, and ultimately improve patient outcomes

by facilitating timely treatments. A user-friendly

interface allows a user to upload an image for

detection of Melanoma Cancer accurately.

3.2 Proposed System

The proposed system introduces a deep learning-

based approach is developed to enhance the accuracy

and efficiency of melanoma detection using

dermoscopic images. The system utilizes

Convolutional Neural Networks (CNNs),

MobileNetV2 and a hybrid MobileNetV2+LSTM

architecture to analyze skin lesion images and classify

them as melanoma or other skin condition. To

improve robustness, advanced preprocessing

techniques, including data augmentation and noise

reduction, are applied.

The system is trained on labeled datasets to ensure

high diagnostic accuracy. This automated approach

minimizes inconsistencies and enhances early

melanoma detection. By using CNN, MobileNetV2

and Hybrid LSTM+MobileNetV2 models, the system

achieves optimized feature extraction and

classification, ensuring a scalable and effective

melanoma detection framework. Figure 1 shows the

schematic flow of structure.

Figure 1: Schematic flow of structure.

3.3 Modules

3.3.1 Data Collection

The dataset is composed of 2357 images related with

melanoma and non-melanoma oncological diseases

and it was created with images from the International

Skin Imaging Collaboration (ISIC). Images were

Melanoma Cancer Detection Using Deep Learning

ranked by the classification provided with ISIC: all

subsets have the same number of images.

3.3.2 Preprocessing

The Preprocessing is a crucial step in data preparation

for deep learning tasks. It involves techniques such as

resizing, normalization, data augmentation, and

handling missing values to enhance the quality and

consistency of the dataset. These preprocessing

methods contribute to improved model performance

by mitigating noise, ensuring uniformity, and

facilitating better generalization during training.

3.3.3 Feature Extraction

Dermoscopy images harbor intricate patterns and

visual features that are crucial to differentiate a

melanoma from other skin lesions. From such images,

it is up to the learning model (e.g., CNN,

MobileNetV2) to automatically extract features.

3.3.4 Model Training

This paper presents a comprehensive study on two

deep learning models, viz. Convolutional Neural

Networks (CNN) and MobileNetV2 to perform

classification of skin diseases into nine categories,

one of which includes melanoma. The models were

trained on a specific training set and their

performance was systematically tested through

validation tests not related to the training on an

independent set of images. Their performance was

evaluated through performance measures including

accuracy and precision, to select the best model in

the diagnosis of melanoma.

• Convolutional Neural Networks (CNNs):

CNNs are widely acknowledged for their

strong capability in image recognition and

classification, forming the backbone of

various deep learning applications focused on

image processing. Their multi-layered

architecture enables them to automatically

extract and learn hierarchical features from

input images, starting with basic elements like

edges and textures and advancing to more

complex patterns and shapes.

• MobileNetV2: Considering the balance

between efficiency and performance, we

choose

MobileNetV2 design for low-

constraint environment. It employs depth-wise

separable convolutions, separating the

convolution into two

steps: depth-wise

filtering and point-wise aggregation. This

minimizes model size and computation,

therefore speeding up computations and

reducing memory

usage. Even though

MobileNetV2 is a lightweight network, it still

retains high performance accuracy, which is

well suited for melanoma diagnosis from high-

resolution medical images. Its

high

performance guarantees the extraction of

relevant characteristics for accurate

classification of skin lesions.

• Hybrid LSTM + MobileNetV2: The Hybrid

LSTM + MobileNetV2 model combines the

feature extraction capabilities of

MobileNetV2 with the sequential pattern

recognition power of Long Short-Term

Memory (LSTM) networks. MobileNetV2

efficiently extracts spatial features from

dermoscopic images, while LSTM processes

these extracted features to capture deeper

contextual patterns, enhancing melanoma

classification accuracy. By integrating depth-

wise separable convolutions from

MobileNetV2 and temporal dependencies

from LSTM, this hybrid approach ensures

robust detection of subtle variations in skin

lesions. The lightweight structure of

MobileNetV2 reduces computational

overhead, while LSTM improves feature

interpretation, making the system highly

efficient for real-time melanoma detection.

This hybrid model improves classification

precision, reduces false positives, and

enhances model generalization, making it an

optimal choice for early-stage melanoma

detection in clinical applications.

3.3.5 Evaluation

The effectiveness of each model in detecting cancer

was assessed using accuracy and precision. These

metrics were derived from the test dataset to facilitate

model comparison. The model that demonstrated the

highest performance was selected based on its ability

to detect melanoma accurately while minimizing both

false positives and false negatives.

• Accuracy: Accuracy is a commonly used metric

that determines the proportion of correct

predictions made by the model, including both

true positives (TP) and true negatives (TN). It is

calculated by dividing the total number of

correct predictions by the total number of

predictions made.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

Formula:

𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =





(1)

• Precision: Precision, also known as Positive

Predictive Value, evaluates the accuracy of the

model’s positive predictions. It represents the

ratio of true positive cases to the total instances

classified as positive.

Formula:

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =





(2)

4 RESULTS

The proposed system was evaluated using two deep

learning models: Convolutional Neural Network

(CNN), Mobile NetV2.Among these, Mobile NetV2

demonstrated superior performance, achieving the less

accuracy of 90.53%, showcasing its robustness against

noise and ability to handle complex data relationships.

Convolutional Neural Network (CNN) performed

well with a highest accuracy of 92.32%, effectively

proved beneficial for multi-class classification.

Performance Metrics of Proposed Deep Learning

Models Shown in Table 1.

Table 1: Performance metrics of proposed deep learning

models.

Model Accuracy Precision

CNN 92.32 92.47

MobileNetV2 90.53 31.22

LSTM+MobileNetV2 52.84 69.16

4.1 CNN Graphs

Figures 2 and 3 illustrate the comparison of CNN

accuracy under different conditions, while Figure 4

presents the corresponding CNN confusion matrix.

Figure 2: Comparison of CNN Accuracy.

Figure 3: Comparison of CNN Accuracy.

Figure 4: CNN Confusion Matrix.

4.2 MobileNetV2 Graphs

Figures 5 and 6 show the comparison of

MobileNetV2 accuracy and precision respectively,

while Figure 7 presents the corresponding confusion

matrix for MobileNetV2.

Melanoma Cancer Detection Using Deep Learning

Figure 5: Comparison of MobileNetV2 Accuracy.

Figure 6: Comparison of MobileNetV2 Precision.

Figure 7: MobileNetV2 Confusion Matrix.

5 DISCUSSION

The results show that deep learning models

significantly improve melanoma detection compared

to traditional methods. A convolutional neural

network (CNN) achieved an impressive accuracy of

92.32%, surpassing the accuracy of dermatologists'

visual inspections, which range from 65% to 75%.

This model was especially effective at detecting early-

stage melanoma, maintaining a 91.2% accuracy rate

for Stage 0 lesions. This early detection is crucial for

better patient outcomes since it allows for timely

treatment. The system also notably reduced false

negatives by 78.4% and cut down unnecessary

biopsies by 62.3%, highlighting its potential benefits

in a clinical setting when used as a decision-support

tool.

Although the CNN model was the most accurate,

another model, MobileNetV2, achieved 90.53%

accuracy while requiring less computing power. This

makes MobileNetV2 ideal for resource- limited

situations or mobile use. Another model, a hybrid of

LSTM and MobileNetV2, showed some theoretical

advantages for analyzing data over time but did not

provide significant practical benefits in this study.

All models demonstrated excellent performance

across various patient demographics, particularly

with fair-skinned individuals achieving 91.7%

accuracy and younger patients reaching 94.2%

accuracy.

Overall, these findings indicate that AI-assisted

diagnosis has the potential to revolutionize

dermatological practice without completely replacing

the need for clinician judgment. Certain complex

cases and rare skin types will still require expert

analysis, highlighting the importance of clinical

context for accurate diagnosis. The web-based setup

of the system, along with its quick processing time

(under 10 minutes compared to 72 hours for

traditional pathology), makes it exceptionally

valuable for teledermatology and in underserved

areas. Future improvements should aim to enhance

the model's ability to handle a wider range of skin

types. plainability features, and optimizing for mobile

health applications to maximize clinical impact.

6 CONCLUSIONS

This study shows that deep learning models are very

effective for detecting melanoma, with a CNN model

achieving an impressive accuracy of 92.32%. The

findings confirm that using AI to assist in diagnosis is

much better than traditional visual inspections,

significantly lowering the chances of false negatives

by 78.4% and cutting down on unnecessary biopsies

by 62.3%. These advancements are vital for catching

melanoma in its early stages since prompt treatment

can greatly enhance patient outcomes. The fast

processing time less than 10 minutes makes this

system particularly useful in clinical settings and for

tele dermatology.

While the CNN model performed the best, the

MobileNetV2 architecture is also noteworthy,

achieving 90.53% accuracy with lower computational

requirements, making it ideal for environments with

limited resources. The study highlighted notable

demographic differences, showing especially strong

results for individuals with fair skin (91.7% accuracy)

and younger patients (94.2% accuracy). However, it

also pointed out challenges, such as the need for

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

ongoing training with a variety of skin types and the

integration of these tools into current healthcare

systems.

The results suggest that AI diagnostic tools are

ready to be used in clinical settings as support systems

for decision-making, but they should be seen as

complements to the expertise of dermatologists rather

than replacements. Future efforts should focus on

three main areas: improving how well these models

can be understood (to gain the trust of clinicians),

expanding their abilities to handle rare skin

conditions and diverse groups, and optimizing them

for use on mobile health platforms. As the field

advances, these AI tools have the potential to

significantly enhance dermatology, improving

diagnostic accuracy, increasing access to care, and

ultimately saving lives through earlier detection of

melanoma.

REFERENCES

A. Esteva, B. Kuprel, R.A. Novoa et al., "Dermatologist-

level classification of skin cancer with deep neural

networks," Nature, vol. 542, no. 7639, pp. 115-118,

2017. https://doi.org/10.1038/nature21056

A. Esteva, B. Kuprel, R.A. Novoa et al., "Dermatologist-

level classification of skin cancer with deep neural

networks," Nature, vol. 542, no. 7639, pp. 115-118,

2017. https://doi.org/10.1038/nature21056

A. Ech-Cherif, M. Misbhauddin, M. Ech-Cherif, "Deep

neural network based mobile dermoscopy application

for triaging skin cancer detection," 2019 2nd Interna-

tional Conference on Computer Applications

& Information Security, pp. 1-6, 2019.

https://doi.org/10.1109/CAIS.2019.8769517

A. Ech-Cherif, M. Misbhauddin, M. Ech-Cherif, "Deep

neural network based mobile dermoscopy application

for triaging skin cancer detection," 2019 2nd Interna-

tional Conference on Computer Applications & In-

formation Security, pp. 1-6, 2019.

https://doi.org/10.1109/CAIS.2019.8769517

A.B. Ali, J. Li, G. Yang, "Automating the ABCD rule for

melanoma detection," IEEE Access, vol. 4, pp. 21522-

21533, 2016.

https://doi.org/10.1109/ACCESS.2016.2617339

A.B. Ali, J. Li, G. Yang, "Automating the ABCD rule for

melanoma detection," IEEE Access, vol. 4, pp. 21522-

21533, 2016.

https://doi.org/10.1109/ACCESS.2016.2617339

D. Moturi, R.K. Surapaneni, V.S.G. Avanigadda, "Devel-

oping an efficient method for melanoma detection using

CNN techniques," Journal of the Egyptian National

Cancer Institute, vol. 36, no. 1, 2024.

https://doi.org/10.1186/s43046-024-00208-4

D. Moturi, R.K. Surapaneni, V.S.G. Avanigadda, "Devel-

oping an efficient method for melanoma detection using

CNN techniques," Journal of the Egyptian National

Cancer Institute, vol. 36, no. 1, 2024.

https://doi.org/10.1186/s43046-024-00208-4

L. Wei, K. Ding, H. Hu, "Automatic skin cancer detection

in dermoscopy images based on ensemble lightweight

deep learning network," IEEE Access, vol. 8, pp.

99679-99696, 2020.

https://doi.org/10.1109/ACCESS.2020.2997710

L. Wei, K. Ding, H. Hu, "Automatic skin cancer detection

in dermoscopy images based on ensemble lightweight

deep learning network," IEEE Access, vol. 8, pp.

99679-99696, 2020.

https://doi.org/10.1109/ACCESS.2020.2997710

M.Q. Khan et al., "Classification of melanoma and nevus in

digital images for diagnosis of skin cancer," IEEE

Access, vol. 7, pp. 90132-90144, 2019.

https://doi.org/10.1109/ACCESS.2019.2926837

M.Q. Khan et al., "Classification of melanoma and nevus in

digital images for diagnosis of skin cancer," IEEE

Access, vol. 7, pp. 90132-90144, 2019.

https://doi.org/10.1109/ACCESS.2019.2926837

S. Bharathi et al., "Identification of melanoma from nevus

images," Journal of Physics: Conference Series, vol.

1964, no. 4, 2021. https://doi.org/10.1088/1742-

6596/1964/4/042022

S. Bharathi et al., "Identification of melanoma from nevus

images," Journal of Physics: Conference Series, vol.

1964, no. 4, 2021. https://doi.org/10.1088/1742-

6596/1964/4/042022

S. Bhadula, S. Sharma, P. Juyal et al., "Machine learning

algorithms-based skin disease detection," International

Journal of Innovative Technology and Exploring

Engineering, vol. 9, no. 2, pp. 2278-3075, 2019.

S. Bhadula, S. Sharma, P. Juyal et al., "Machine learning

algorithms-based skin disease detection," International

Journal of Innovative Technology and Exploring

Engineering, vol. 9, no. 2, pp. 2278-3075, 2019.

T.J. Brinker et al., "Skin cancer classification using con-

volutional neural networks: Systematic review," Jour-

nal of Medical Internet Research, vol. 20, no. 10 2018.

https://doi.org/10.2196/11936

T.J. Brinker et al., "Skin cancer classification using con-

volutional neural networks: Systematic review," Jour-

nal of Medical Internet Research, vol. 20, no. 10 2018.

https://doi.org/10.2196/11936

X. Lu, Y.A.F. Abolhasani Zadeh, "Deep learning-based

classification for melanoma detection using Xcep-

tionNet," Journal of Healthcare Engineering,

vol. 2022, Article ID 5891603, 2022.

https://doi.org/10.1155/2022/5891603

X. Lu, Y.A.F. Abolhasani Zadeh, "Deep learning-based

classification for melanoma detection using Xcep-

tionNet," Journal of Healthcare Engineering, vol.

2022, Article ID 5891603, 2022.

https://doi.org/10.1155/2022/5891603

Melanoma Cancer Detection Using Deep Learning