Early Melanoma Detection Using Deep Learning: A Comprehensive Review

Abstract

Melanoma is one of the most aggressive and life-threatening forms of skin cancer, posing a major challenge to global healthcare due to its rapid progression and high mortality rates when diagnosis is delayed. Conventional diagnostic methods primarily depend on clinical observation and dermatologist expertise, which can often be subjective and inconsistent. In recent years, deep learning—particularly Convolutional Neural Networks (CNNs)—has shown strong potential to enhance diagnostic accuracy by automatically identifying subtle variations in dermoscopic images that may not be easily recognized by the human eye. This study presents a CNN-based framework for the early detection of melanoma using transfer learning with the ResNet50 architecture. The research employs the HAM10000 dataset, which contains over 10,000 dermoscopic images representing seven different skin lesion categories. Prior to training, the dataset undergoes normalization and data augmentation to improve image quality and model generalization. The ResNet50 network is fine-tuned by replacing its fully connected layers with custom dense, dropout, and sigmoid layers, optimized using the Adam optimizer and binary cross-entropy loss function. The proposed approach delivers a reliable, non-invasive, and efficient diagnostic support system to assist dermatologists in early melanoma recognition. By combining transfer learning and augmentation, the model improves classification accuracy and reduces dependency on manual feature extraction. Overall, this work highlights the promise of deep learning in dermatological diagnostics and lays the groundwork for future research on explainable models and clinical validation in real-world medical settings.

Introduction

Melanoma, the deadliest form of skin cancer, spreads rapidly and requires early detection for effective treatment. Traditional diagnosis relies on dermatologists’ visual inspection through dermoscopy, which is time-consuming and prone to human error due to variations in color, shape, and texture of skin lesions. With advances in deep learning, particularly Convolutional Neural Networks (CNNs), automated melanoma detection has emerged as a reliable, non-invasive diagnostic approach.

This study proposes a CNN-based deep learning system using transfer learning with the ResNet50 architecture to classify melanoma and non-melanoma lesions. The model is trained and validated on the HAM10000 dataset, which includes 10,015 dermoscopic images from seven skin lesion categories. The goal is to assist clinicians by providing an efficient, accurate, and automated tool for melanoma diagnosis.

Literature Review & Research Gaps

Recent research has demonstrated strong CNN performance for skin cancer detection:

• Jojoa Acosta et al. (2021) achieved 90.4% accuracy using Mask R-CNN + ResNet152 but ignored lesion progression over time.

• Garcia et al. (2021) applied meta-learning for limited data, but scalability to larger datasets remains untested.

• Alheejawi et al. (2021) achieved 97.7% segmentation accuracy on histopathological images; adaptation to dermoscopic data is lacking.

• Yu et al. (2022) introduced temporal lesion analysis, outperforming clinicians but requiring validation on public datasets.

• Musthafa et al. (2024) and Moturi et al. (2024) optimized CNNs (ResNet50, MobileNetV2) for HAM10000, achieving up to 97.78% accuracy, but focused on static images.

• Esmaeili et al. (2025) and Azhari et al. (2025) proposed advanced CNN frameworks (multi-stream and two-stage models) with accuracies nearing 99–100%, though they risk overfitting and lack external clinical validation.

Common research gaps include: limited use of temporal lesion data, lack of generalization testing, and reliance on static, benchmark datasets.

Proposed Methodology

The proposed system follows four main stages:

1. Dataset Preparation: Uses the HAM10000 dataset, resized to 224×224 pixels, normalized, and augmented (rotation, flipping, brightness, zoom) to improve generalization.

2. Model Design: A fine-tuned ResNet50 pretrained on ImageNet, modified for binary classification (melanoma vs. non-melanoma) with:

   • Global Average Pooling, Dense (128 ReLU), Dropout, and Sigmoid layers.

3. Training Strategy:

   • 70% training, 20% validation, 10% testing.

   • Adam optimizer, learning rate = 0.001, binary cross-entropy loss.

   • Early stopping and checkpointing prevent overfitting.

4. Evaluation Metrics: Accuracy, Precision, Recall, F1-score, ROC–AUC, and Confusion Matrix for comprehensive performance analysis.

Conclusion

Early and accurate detection of melanoma remains a critical factor in improving patient survival rates and reducing the global burden of skin cancer. This review highlights the increasing role of deep learning, particularly Convolutional Neural Networks (CNNs), as a dependable diagnostic aid in dermatology. By automatically identifying intricate visual patterns such as color, texture, and lesion borders, CNN-based models address many of the limitations associated with manual diagnosis and inter-observer variability. Among existing architectures, transfer learning techniques using advanced networks like ResNet50 and hybrid CNN models have consistently demonstrated high accuracy and robustness on benchmark datasets such as HAM10000 and ISIC. Despite these promising advances, certain research gaps remain. Most existing systems rely solely on static dermoscopic images and do not account for temporal lesion changes—an important factor for early-stage melanoma detection. Furthermore, while many models show strong performance on benchmark datasets, their effectiveness in real-world clinical settings requires further validation. Expanding dataset diversity, incorporating temporal and multimodal learning frameworks, and improving model generalization could significantly enhance diagnostic reliability. In summary, deep learning continues to transform the field of melanoma detection by offering efficient, scalable, and non-invasive diagnostic solutions. The integration of these intelligent systems into clinical workflows can provide valuable support to dermatologists, leading to earlier intervention and improved patient outcomes. Future research should focus on enhancing model transparency, cross-dataset validation, and clinical decision-support integration to ensure the safe, ethical, and effective adoption of AI-driven diagnostic systems in healthcare practice.

References

[1] M. M. Rahman, A. Chowdhury, and S. Hossain, “A hybrid CNN-based approach for early detection of skin cancer,” IEEE Xplore, document no. 9630443, 2021. [2] A. P. Singh and B. K. Balabantaray, “Skin lesion classification using hybrid deep learning framework,” IEEE Xplore, document no. 9570281, 2021. [3] F. Haghshenas, “Comparative study of deep learning models in melanoma detection,” in Artificial Neural Networks in Pattern Recognition, Springer, Cham, pp. 121–131, 2024. [4] N. Behera, A. P. Singh, J. K. Rout, and B. K. Balabantaray, “Melanoma skin cancer detection using deep learning-based lesion segmentation,” International Journal of Information Technology, vol. 16, no. 8, pp. 3729–3744, 2024. [5] M. M. Musthafa, R. R. Kumar, and M. S. Basha, “Enhanced skin cancer diagnosis using optimized CNN architecture,” International Journal of Information Technology, vol. 16, no. 8, pp. 3729–3744, 2024. [6] D. Moturi, R. K. Surapaneni, and V. S. G. Avanigadda, “Developing an efficient method for melanoma detection using CNN techniques,” Journal of the Egyptian National Cancer Institute, vol. 36, no. 1, pp. 1–10, Feb. 2024. [7] V. Esmaeili, S. M. A. Mirmohseni, and P. Rad, “Automatic melanoma detection using an optimized five-stream convolutional neural network,” Scientific Reports, vol. 15, no. 1, pp. 1–10, 2025. [8] M. Tschandl, C. Rosendahl, and H. Kittler, “The HAM10000 dataset: A large collection of multi-source dermatoscopic images of common pigmented skin lesions,” Scientific Data, vol. 5, article 180161, pp. 1–9, 2018. [9] S. Ramesh, S. B. Nandhini, and K. G. Rajesh, “Melanoma skin cancer detection using deep learning,” Journal of Engineering and Clinical Investigations, vol. 5, no. 2, pp. 112–119, 2024. [10] A. A. Azhari, N. Yudistira, A. W. Widodo, and Y. Yagi, “Two-stage CNN with weakly supervised segmentation for skin lesion classification,” Multimedia Tools and Applications, vol. 84, no. 6, pp. 16217–16235, 2025. [11] M. F. Jojoa Acosta, L. Y. Caballero Tovar, M. B. Garcia-Zapirain, and W. S. Percybrooks, “Melanoma diagnosis using deep learning techniques on dermatoscopic images,” BMC Medical Imaging, vol. 24, no. 1, pp. 1–11, 2024. [12] Z. Yu, H. Jiang, L. Zhang, and Y. Ding, “Early melanoma diagnosis with sequential dermatoscopic images,” Computers in Biology and Medicine, vol. 145, pp. 105460–105472, 2022. [13] S. I. Garcia, “Meta-learning for skin cancer detection using deep learning techniques,” Biomedical Signal Processing and Control, vol. 68, pp. 102728–102737, 2021.

Copyright

Copyright © 2025 Tejas P Abnave, Shashikant V. Athawale, Snehal P. Kamale, Harshwardhan M. Bhorde, Neha A. Kamlesh. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.