Ensemble-Based Automated Skin Disease Classification Using Deep Convolutional Neural Networks

Abstract

Skin disorders are common across all age groups and can lead to serious complications if not identified early. Conventional diagnosis methods rely on dermatologists\' visual examination, which can be slow, subjective, and prone to error, particularly due to the visual similarity between different lesion types. These challenges are further amplified in areas with limited access to specialized healthcare. To address this issue, we propose a deep learning-based ensemble model for automatic skin disease classification. The Employs an ensemble of advanced deep convolutional models, namely Xception, InceptionV3, and ResNet50 each fine-tuned on the HAM10000 dermoscopic image dataset, which includes seven distinct skin lesion categories. To enhance the model’s performance and handle class imbalance, we apply techniques such as image augmentation, class weighting, random oversampling, and test-time augmentation (TTA). Final predictions are obtained through weighted soft voting across the ensemble. Additionally, Grad-CAM is employed to generate visual explanations of the model’s predictions, promoting transparency and aiding clinical interpretability. The proposed method achieves an accuracy of 97.22%, offering a robust and scalable solution that could support early diagnosis and improve dermatological services, especially in regions with limited healthcare infrastructure.

Introduction

Skin diseases are prevalent across all age groups, and accurate, timely diagnosis is essential. Traditional diagnosis by dermatologists through visual inspection can be subjective, time-consuming, and prone to errors due to visual similarity among different skin lesions. Deep learning, especially Convolutional Neural Networks (CNNs), offers a powerful solution by automating feature extraction and classification from dermoscopic images.

This study proposes an ensemble-based CNN classification system that integrates Xception, InceptionV3, and ResNet50 models. The system is trained on the HAM10000 dataset containing over 10,000 labeled dermoscopic images from seven lesion categories. Key techniques used to improve performance include data augmentation, oversampling, class weighting, and Test-Time Augmentation (TTA). Grad-CAM is used for visual interpretability, highlighting areas of images that influenced model predictions.

Model Architecture and Process:

• Preprocessing: Images resized to 299×299, pixel normalization applied.

• Augmentation: Includes flips, rotations, zoom, and brightness changes to enhance generalization.

• Model Training: Fine-tuned pretrained CNNs using transfer learning with techniques to handle class imbalance.

• Ensemble Prediction: Weighted soft voting (Xception: 0.4, InceptionV3: 0.3, ResNet50: 0.3) is used for final prediction.

• Interpretability: Grad-CAM visualizations provide explainable outputs to aid clinical trust.

Model Contributions:

• Xception: Efficient feature learning through depthwise separable convolutions.

• InceptionV3: Multi-scale feature extraction using parallel convolutions.

• ResNet50: Residual connections allow deep learning with reduced vanishing gradient issues.

Results:

• The ensemble model outperforms individual models and prior approaches, achieving 97.15% accuracy on the HAM10000 dataset.

• In training:

  • Xception and InceptionV3 showed consistent performance with minimal overfitting.

  • ResNet50 exhibited overfitting signs despite high training accuracy.

• The system offers both high diagnostic accuracy and clinical interpretability.

Related Work Comparison:

• Prior models achieved accuracies between 80% to 96.89%, often limited by shallow networks or poor handling of class imbalance.

• The proposed method surpasses these results through model ensembling, robust data strategies, and explainable AI techniques.

Applications and Future Potential:

• Ideal for clinical and telemedicine settings.

• Supports real-time diagnosis, early detection, and prioritization of serious conditions.

• Particularly useful in regions lacking medical experts.

• With ongoing deep learning advancements, ensemble models like this are expected to become integral in automated skin disease screening and diagnosis.

Conclusion

This study presents a deep learning–based ensemble framework for the automated classification of skin diseases, integrating three well-established CNN architectures: Xception, InceptionV3, and ResNet50. By fine-tuning these models on the HAM10000 dermoscopic image dataset and employing techniques such as data augmentation, oversampling, class weighting, and test-time augmentation (TTA), the system achieved a notable classification accuracy of 97.08%. This performance exceeds that of individual models and demonstrates the effectiveness of ensemble learning in handling diverse lesion types. To enhance interpretability, Grad-CAM was used to generate heatmaps that visualize the key regions influencing each prediction. This improves the model’s transparency, helping clinicians understand the reasoning behind the automated diagnosis. As a result, the system not only performs accurate classifications but also offers insights that build trust in AI-assisted dermatological tools. Despite its strong performance, the framework is limited by its dependence on a single dataset, which may affect its generalizability across different populations or imaging conditions. Future research should focus on incorporating larger, more diverse datasets to improve robustness. Exploring advanced architectures such as Vision Transformers or hybrid models, and refining the ensemble strategy could further boost performance. Additionally, developing an intuitive, clinician-friendly interface and integrating patient metadata or multimodal inputs (e.g., clinical images, patient history) may enhance diagnostic precision and usability in real-world healthcare environments.

References

[1] P. R. Priya and K. Ramesh, “Skin Disease Detection Using Deep Learning,” Int. J. for Research in Applied Science and Engineering Technology (IJRASET), vol. 11, no. 4, pp. 182–186, 2023. [2] S. Karthik and B. Nandhini, “CNN-Based Classification of Skin Lesions,” Int. J. of Engineering Research & Technology (IJERT), vol. 13, no. 2, pp. 105–110, 2024. [3] M. Jaya and A. Srikanth, “Deep Learning for Skin Lesion Analysis,” Int. J. of Trend in Scientific Research and Development (IJTSRD), vol. 7, no. 1, pp. 154–160, 2023. [4] V. Lakshmi and G. Preethi, “Automated Skin Disease Classification using CNN,” Int. J. of Scientific and Engineering Research (IJSER), vol. 14, no. 3, pp. 72–77, 2023. [5] A. Keerthana and R. Sudha, “Early Detection of Skin Cancer using CNN,” Int. Research J. of Engineering and Technology (IRJET), vol. 11, no. 1, pp. 921–925, 2024. [6] B. Pooja and R. Arul, “Deep CNN Approach for Skin Disease Classification,” Int. J. of Creative Research Thoughts (IJCRT), vol. 11, no. 2, pp. 432–437, 2023. [7] S. Anitha and M. Rajesh, “Skin Lesion Detection using Hybrid CNN,” Int. J. of Engineering and Advanced Technology (IJEAT), vol. 13, no. 2, pp. 188–192, 2024. [8] K. Swetha and A. Devi, “Deep Learning-Based Skin Disease Recognition,” Int. J. of Research in Engineering and Advanced Management (IJREAM), vol. 9, no. 3, pp. 299–303, 2023. [9] G. Meghana and D. Ramesh, “Skin Cancer Classification using Deep CNN,” Int. J. of Modern Engineering & Technology in Management Research (IJMETMR), vol. 9, no. 4, pp. 117–123, 2024. [10] S. Thirunavukkarasu, E. Paul, R. Singaravelan, and V. S. Srinivasan, “Skin disease detection using CNN,” AIP Conf. Proc., vol. 2857, art. 020055, 2023. [11] S. S. Tantry, T. S. Lal, and S. Rangaswamy, “Skin disease classification using deep learning (InceptionV3 & VGG16),” Natl. Acad. Sci. Lett., 2024. [12] M. Pyingkodi, G. Chandrasekaran, S. Palarimath, V. P., R. Rahul, and V. Vibin, “Exploring deep learning algorithms for skin disease classification using skin lesion images,” in Proc. 3rd Int. Conf. Optimization Techniques Eng. (ICOFE), 2024. [13] S. Innani, P. Dutande, B. Baheti, U. Baid, and S. Talbar, “Deep learning based novel cascaded approach for skin lesion analysis,” arXiv preprint arXiv:2301.00000, 2023. [14] D. Boddupelli, B. Chandana, Sandhyarani, and N. Bhaskar, “Deep learning based skin disease detection using convolutional neural networks (CNN),” Unpublished Indian Conference, 2023. [15] D. Lalrinawma, V. D. A. Kumar, and R. Lalchhanhima, “Image processing and deep learning for classification of skin lesion image,” in Proc. 5th Int. Conf. Smart Electron. Commun. (ICOSEC), Trichy, India, 2024. [16] M. K. Ojha, D. Rose, A. D. Sharma, and S. M. Bency, “Skin Disease Detection and Classification,” Int. J. of Engineering Research & Technology (IJERT), vol. 10, no. 5, pp. 61–64, 2021. [17] T. Parashar, K. Joshi, R. R. N., D. Verma, N. Kumar, and K. S. Krishna, “Skin Disease Detection using Deep Learning,” Int. J. of Health Sciences, vol. 6, no. S1, pp. 6611–6621, 2022. [18] T. Shanthi, R. S. Sabeenian, and R. Anand, “Automatic diagnosis of skin diseases using convolution neural network,” Materials Today: Proceedings, vol. 33, pp. 4560–4567, 2020. [19] A. Murugan, S. A. H. Nair, A. A. P. Preethi, and K. P. S. Kumar, “Diagnosis of skin cancer using machine learning techniques,” Materials Today: Proceedings, vol. 62, pp. 1488–1494, 2022.

Copyright

Copyright © 2025 Venkata Sravani Kaviti, Dr. Haritha Dasari. 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.