An Enhanced CNN-Based Framework for Early Skin Cancer

Detection
1st N. Vijayakumar 2nd Benlo Varghese E S
Assistant Professor, Department of Computer 3rd Arun Raj M
Computer Science and Science and Department of Computer
Engineering, Pollachi Engineering, Science and
Institute of Engineering and Pollachi Institute of Engineering,
Technology, Engineering and Technology, Pollachi Institute of
Pollachi,Coimbatore,India Engineering and Technology,
Pollachi, Coimbatore,India.
nvijay3988@[Link] benloformation@[Link] Pollachi,Coimbatore,India
thzarunraja@[Link]
4th [Link] 5th G Mohanraj
Department of Computer Department of Computer 6th Pavin Balaji R A
Science and Science and Department of Computer
Engineering, Engineering, Science and
Pollachi Institute of Pollachi Institute of Engineering,
Engineering and Technology, Engineering and Technology, Pollachi Institute of
Pollachi,Coimbatore,India Pollachi,Coimbatore,India Engineering and Technology,
suseela225k@[Link] mohanrajgandhi2003@gmail. Pollachi,Coimbatore,India
com pavinbalaji18@[Link]

Abstract— Skin cancer is one of the most Malignant vs. Benign Classification, Diagnostic
common and potentially life-threatening forms Accuracy
of cancer, with early detection being crucial for
effective treatment and improved survival I. INTRODUCTION
rates. In this study, we propose a deep Skin cancer is a growing global health concern,
learning-based approach using Convolutional with millions of new cases reported annually. It is
Neural Networks (CNNs) to predict skin cancer broadly categorized into melanoma and non-
from dermoscopic images. The proposed model melanoma skin cancers, with melanoma being the
leverages a robust CNN architecture to classify deadliest form due to its high metastatic potential.
skin lesions into malignant or benign Early and accurate detection is critical to
categories. The dataset, sourced from publicly improving patient outcomes and reducing
available repositories such as the ISIC Archive, mortality rates. However, traditional diagnostic
was preprocessed through techniques like data methods, such as visual inspection and biopsy, are
augmentation and normalization to enhance often time-consuming, invasive, and dependent on
model performance. the expertise of dermatologists.
Recent advancements in artificial intelligence
The model achieved promising results, (AI) and deep learning have shown great promise
demonstrating its potential in clinical in addressing these challenges. Among these,
applications. Our analysis highlights the Convolutional Neural Networks (CNNs) have
effectiveness of deep learning in automating emerged as a powerful tool for image-based
and improving diagnostic accuracy while classification tasks, including medical imaging.
reducing the need for invasive procedures. CNNs excel at learning hierarchical features from
Future work aims to extend the model for raw image data, making them ideal for identifying
multi-class classification and integrate it into complex patterns in dermoscopic images of skin
real-time diagnostic tools. lesions.
This study focuses on leveraging CNNs to
Keywords-
develop an automated system for skin cancer
CNN,DeepLearning,DermoscopicImages,
prediction. By analyzing dermoscopic images, the
proposed model aims to classify skin lesions as
malignant or benign with high accuracy, offering a classified them using SVMs. However, these
non-invasive and efficient diagnostic alternative. methods faced limitations in capturing complex
The key contributions of this work include: patterns, leading to suboptimal diagnostic
performance.
1. Designing a robust CNN architecture
tailored for skin lesion classification. With the rise of deep learning, particularly
2. Employing advanced preprocessing Convolutional Neural Networks (CNNs),
techniques, such as data augmentation, to researchers have achieved significant
enhance model performance. breakthroughs in medical image analysis. Esteva
3. Evaluating the model using standard et al. [6] demonstrated that a deep CNN trained on
metrics like accuracy, sensitivity, and over 129,000 images could achieve dermatologist-
specificity to ensure clinical relevance. level performance in identifying melanoma and
other skin diseases. Similarly, Codella et al. [7]
The remainder of this paper is organized as presented an ensemble framework combining
follows: Section II discusses related work in skin
CNNs with traditional features, which achieved
cancer prediction using CNNs. Section III
describes the proposed methodology, including high diagnostic accuracy on the ISIC Challenge
dataset selection, preprocessing, and model dataset.
architecture. Section IV presents the experimental
results, while Section . Transfer learning has proven effective in
addressing the challenges of limited labeled
While CNNs have shown great potential in skin medical datasets. Studies like Menegola et al. [8]
cancer detection, challenges such as variability in and Kawahara et al. [9] utilized pretrained models
image quality, class imbalance, and limited data such as ResNet and Inception, fine-tuned on
availability still hinder their broader application.
smaller dermoscopic datasets, to enhance
These issues can lead to reduced accuracy and
generalizability, especially when models are tested performance. The use of ensemble methods, as
on diverse datasets. highlighted by Hosny et al. [10], has also shown
improvements in sensitivity and specificity for
Transfer learning has proven particularly skin lesion classification.
effective in cases with limited labeled data,
allowing pre-trained models like ResNet and More recently, attention mechanisms and hybrid
Inception to be fine-tuned for specific tasks [1][2]. models have been explored to improve
Data augmentation methods, such as image
classification performance. Li et al. [11]
rotation and flipping, have further enhanced model
introduced an attention-guided CNN that
robustness by creating diverse training datasets [3].
However, challenges related to interpretability and dynamically focuses on lesion-specific regions,
real-world generalizability continue to hinder the improving interpretability. Furthermore, Ayan and
widespread adoption of these models in clinical Erbay [12] experimented with a hybrid deep
workflows [4]. learning model combining CNNs with recurrent
neural networks (RNNs) to analyze temporal data
II. LITERATURE REVIEW in lesion progression.

The application of machine learning and deep Despite these advancements, challenges remain in
learning in skin cancer detection has been improving model generalizability, handling class
extensively studied in recent years, yielding imbalance in datasets, and enhancing
promising results. This section highlights notable interpretability. Most studies focus on binary
contributions in this domain and identifies the classification (benign vs. malignant), while multi-
gaps that our work aims to address. class classification, which is critical for
differentiating subtypes of skin cancer, remains
Early approaches relied on traditional machine underexplored. Additionally, the lack of real-time
learning algorithms, such as Support Vector clinical integration limits the practical utility of
Machines (SVMs) and Random Forests, applied to these systems.
handcrafted features like color, texture, and shape.
For example, Barata et al. [5] extracted color and This study builds on prior work by leveraging a
texture features from dermoscopic images and robust CNN architecture tailored for binary
classification of skin lesions. Advanced kernels/filters, pooling layers for reducing spatial
preprocessing techniques, including data dimensions and computational complexity via
augmentation, are employed to address class max-pooling, fully connected layers to flatten
imbalance. The performance of the model is feature maps and pass them through dense layers
evaluated using clinically relevant metrics such as for classification, and an output layer utilizing a
accuracy, sensitivity, and specificity, providing softmax activation function for multiclass
insights into its potential for real-world classification or sigmoid for binary classification.
applications. Pre-trained architectures such as VGG16, ResNet,
and InceptionNet were explored for transfer
III. METHODOLOGY learning to leverage their pre-existing feature
extraction capabilities.
Dataset Collection:To conduct this research,
dermoscopic images of skin lesions were obtained
Training and Hyperparameter Optimization:The
from publicly available datasets, such as the ISIC
model was trained on the dataset with the
(International Skin Imaging Collaboration)
following configurations: Binary Cross-Entropy
Archive. This dataset comprises labeled images
loss for binary classification or Categorical Cross-
categorized into classes such as benign and
Entropy for multiclass classification [13], Adam
malignant. The dataset was chosen due to its
optimizer for efficient gradient descent [14], a
quality, diversity, and relevance to the study.
dynamic learning rate adjusted using a learning
rate scheduler [15], a batch size of 32, and training
for 50 epochs to ensure convergence.
Additionally, 20% of the dataset was reserved for
validation to monitor model performance during
training.

Evaluation Metrics:The performance of the model

was assessed using a variety of evaluation metrics
to provide a comprehensive understanding of its
effectiveness. Accuracy was used as the primary
measure to determine the overall correctness of
the model's predictions. However, since the
dataset was imbalanced, Precision, Recall, and F1-
Score were also considered to evaluate the model's
ability to correctly identify instances from both
the majority and minority classes. These metrics
are particularly useful in scenarios where one
Figure [Link] of Nine Disease class is significantly more frequent than the other,
as they provide a more detailed view of the
Data Processing: The collected dermoscopic model's performance
images were resized to a uniform dimension of
224x224 pixels, normalized to a range of [0, 1], Implementation Environment:The model was
and augmented using techniques such as rotation, implemented using Python, leveraging various
flipping, zooming, and brightness adjustments to libraries and tools to ensure efficient development
enhance the model's generalization capability. and execution. TensorFlow/Keras was utilized for
Additionally, class imbalance issues were building and training the Convolutional Neural
addressed by employing oversampling or Network (CNN), while NumPy and Pandas were
undersampling techniques to ensure a balanced employed for data manipulation and preprocessing
dataset for training the model. tasks. To visualize the results and gain insights
into the model's performance, Matplotlib and
Model Architecture:The Convolutional Neural Seaborn were used for data visualization.
Network (CNN) served as the core predictive Additionally, Google Colab with GPU support
model in this research, consisting of an input layer was leveraged to enhance computational
for preprocessing images, convolutional layers for
extracting spatial and hierarchical features with
efficiency, enabling faster model training and
experimentation.

Deployment:After training and evaluation, the

final model was deployed in a user-friendly
interface for real-world application. The interface
allows users to upload dermoscopic images and
receive predictions about the likelihood of
malignancy.
IV. RESULTS AND ANALYSIS

Model Performance

Training Loss: The model achieved a training loss

of 1.0578, indicating effective learning as the Figure [Link] and validation Accuracy
error between predicted and actual outcomes
decreased during training. This reflects the
Best Training Epoch:
model's ability to capture patterns in the data.
The epoch with the highest training accuracy was
Validation Loss: The model achieved a validation identified, marking the point at which the model
loss of 2.1410, suggesting room for improvement achieved its optimal performance on the training
in generalizing to unseen data. This highlights the data. This epoch signifies the moment when the
need for further fine-tuning or regularization to model successfully generalized the learned
enhance model performance on validation features and showed the best fit for the training set.
datasets. Best Validation Epoch:
Training Accuracy: The model achieved a training The epoch with the lowest validation loss was
accuracy of 93.75%, indicating that it effectively also noted, representing the point where the model
learned the patterns in the training dataset and exhibited its best generalization to unseen data.
performed well on the provided data.. This epoch indicates the model's ability to
maintain a low error rate on the validation set,
confirming its robustness and minimizing the risk
of overfitting
Evaluation Metrics:
The model achieved a train loss of 1.0579 and a
train accuracy of 93.75%. The validation loss was
2.1410, with a validation accuracy of 59.32%. For
the test set, the model resulted in a test loss of
2.5699 and a test accuracy of 45.76%.

Figure [Link] and Validation Loss

Figure [Link] of Model Metrics
Validation Accuracy: The model achieved a
validation accuracy of 59.32%, suggesting Confusion Matrix:
moderate performance on unseen data. This result
indicates potential room for improvement, 1. Accuracy
possibly through hyperparameter tuning,  Accuracy = (TP + TN) / (TP + TN + FP +
regularization, or data augmentation. FN)
  TP = 6 + 3 + 6 + 3 + 2 + 10 = 30 Squamous Cell Carcinoma, and Actinic Keratosis
 TN = 8 + 9 + 3 + 1 + 1 + 1 + 2 + 2 = 27 had a precision, recall, and F1-score of 0.00,
indicating poor performance on these classes. The
 FP = 10 + 0 + 1 + 1 + 6 + 3 + 3 + 2 = 26 model's overall accuracy was 46%, with the macro
 FN = 0 + 0 + 1 + 1 + 0 + 0 + 2 + 0 = 4 average precision of 0.38, recall of 0.48, and F1-
score of 0.39, and the weighted average precision
 Accuracy = (30 + 27) / (30 + 27 + 26 + 4) of 0.38, recall of 0.46, and F1-score of 0.38.
= 57 / 87 ≈ 0.655
2. Precision (for Melanoma)
 Precision (Melanoma) = TP_Melanoma /
(TP_Melanoma + FP_Melanoma)
 TP_Melanoma = 6
 FP_Melanoma = 1
 Precision (Melanoma) = 6 / (6 + 1) = 6/7 ≈
0.857
3. Recall (for Melanoma)
 Recall (Melanoma) = TP_Melanoma / Figure [Link] Report
(TP_Melanoma + FN_Melanoma)
 TP_Melanoma = 6 V. CONCLUSION
 FN_Melanoma = 1
This research used a Convolutional Neural
 Recall (Melanoma) = 6 / (6 + 1) = 6/7 ≈ Network (CNN) to classify different types of skin
0.857 cancer. The model showed good performance
4. F1-Score (for Melanoma) across all classes, with the ability to identify
various skin conditions. However, there is still
 F1-Score (Melanoma) = 2 * (Precision *
room for improvement in precision and recall for
Recall) / (Precision + Recall)
some categories.
 F1-Score (Melanoma) = 2 * (0.857 *
0.857) / (0.857 + 0.857) = 0.857 Future work will focus on enhancing the model by
expanding the dataset, fine-tuning
hyperparameters, and applying techniques such as
data augmentation. The aim is to improve the
model's accuracy and robustness, contributing to
the development of reliable tools for early skin
cancer detection to assist healthcare professionals
in making informed diagnoses.

REFERENCE
[1].Kawahara, J., Daneshvar, S., Argenziano,
G., & Hamarneh, G. (2018). Seven-point
checklist and skin lesion classification
using multitask multimodal neural nets.
[2].Hosny, A., Parmar, C., Quackenbush, J., et
Figure [Link] matrix al. (2020). Artificial intelligence in
radiology.
Classification Report: [3].Li, Z., Wang, Y., Yu, S., et al. (2021).
Attention-based deep learning system for
The model's performance across different classes skin lesion diagnosis.
was varied, with Basal Cell Carcinoma achieving [4].Ayan, E., & Erbay, H. (2019). Diagnosis
the highest precision (0.82), recall (1.00), and F1- of melanoma skin cancer using hybrid
score (0.90), while Seborrheic Keratosis, deep learning models.
[5].Barata, C., Celebi, M. E., & Marques, J. S. [13]. Chollet, F. (2017). Deep Learning
(2013). Improving dermoscopy image with Python. Manning Publications.
classification using color and texture [14]. Kingma, D.P., & Ba, J. (2014).
features. IEEE Transactions on Adam: A method for stochastic
Biomedical Engineering. optimization. arXiv preprint
[6].Esteva, A., Kuprel, B., Novoa, R. A., et al. arXiv:1412.6980.
(2017). Dermatologist-level classification [15]. Smith, L. N. (2017). Cyclical
of skin cancer with deep neural networks. Learning Rates for Training Neural
Nature. Networks. arXiv preprint
[7].Codella, N., Cai, J., Abedini, M., et al. arXiv:1506.01186.
(2018). Skin lesion analysis toward
melanoma detection: A challenge at the
2017 ISIC Dataset. IEEE Transactions on
Medical Imaging.
[8].Menegola, A., Tavares, J., Fornaciali, M.,
et al. (2018). RECOD Titans at ISIC
Challenge 2018. arXiv preprint
arXiv:1809.09960.
[9].Kawahara, J., Daneshvar, S., Argenziano,
G., & Hamarneh, G. (2018). Seven-point
checklist and skin lesion classification
using multitask multimodal neural nets.
IEEE Journal of Biomedical and Health
Informatics.

[10]. Hosny, A., Parmar, C.,

Quackenbush, J., et al. (2020). Artificial
intelligence in radiology. Nature Reviews
Cancer.
[11]. Li, Z., Wang, Y., Yu, S., et al.
(2021). Attention-based deep learning
system for skin lesion diagnosis.
Computer Methods and Programs in
Biomedicine.
[12]. Ayan, E., & Erbay, H. (2019).
Diagnosis of melanoma skin cancer using
hybrid deep learning models. Medical
Hypotheses.