Skin Disease Detection Using Convolutional Neural Networks

Abstract

Skin diseases continue to be one of the most widely diagnosed medical concerns globally, affecting both developed and developing regions at scale. These dermatological conditions, which range from fungal infections and viral lesions to chronic autoimmune disorders, demand timely and precise identification to ensure effective treatment. Unfortunately, diagnosis often hinges on access to skilled dermatologists and clinical resources, both of which may be limited in rural or underserved areas. Manual diagnosis is not only time-consuming but prone to errors due to visual overlap across diseases. To address this challenge, our research presents a robust deep learning-based classification system that uses Convolutional Neural Networks (CNNs) to automatically detect a wide range of skin diseases from image data. We compiled a diverse and high-resolution custom dataset consisting of over 8,000 annotated images covering various dermatological conditions including Chickenpox, Shingles, Psoriasis, Nail Fungus, Cutaneous Larvae Migrans, Impetigo, and several others. We trained a ResNet50-based model, which leveraged transfer learning and advanced preprocessing strategies to enhance classification performance. Our model attained over 92% accuracy, demonstrating high generalization capability. In this paper, we explore the dataset preparation process, CNN design, training methodology, and evaluation metrics in-depth. Furthermore, we integrate state-of-the-art explainability techniques and discuss practical deployment strategies to ensure real-world applicability. The results strongly support CNNs as reliable tools for aiding dermatological diagnostics in both clinical and mobile telemedicine applications.

Introduction

The paper presents a deep learning-based approach for multi-class skin disease classification using a custom-built dataset and a ResNet50 convolutional neural network (CNN) architecture. The study aims to provide an accurate, scalable, and accessible diagnostic tool for skin diseases, addressing issues such as misdiagnosis, limited dermatological expertise, and lack of infrastructure—especially in resource-constrained areas.

Key Highlights:

Background & Motivation:

• Human skin is exposed to numerous harmful agents, making it vulnerable to a wide range of diseases (e.g., Chickenpox, Psoriasis, Impetigo).

• Traditional diagnostic methods require dermatologists and lab infrastructure, which are not always available.

• Deep learning, particularly CNNs, offers a solution by automating visual pattern recognition in medical images.

Literature Review:

• Early AI models (e.g., Esteva et al., 2017) showed dermatologist-level performance in melanoma detection.

• Recent works (e.g., Gururaj et al., 2023) explored hybrid CNN models like DeepSkin for cancer classification.

• However, existing studies often focus on binary cancer diagnosis, use limited disease types, and datasets lack ethnic and environmental diversity.

• This research fills the gap by focusing on non-cancerous diseases and a diverse, custom dataset.

Dataset Preparation:

• 8,000 images from varied sources, across 7 disease classes:

  • Chickenpox, Shingles, Psoriasis, Nail Fungus, Cutaneous Larvae Migrans, Impetigo, Tinea Corporis

• Preprocessing: Resizing (224x224), normalization, noise filtering, and color enhancement.

• Augmentation: Rotation, zoom, flips, brightness changes, and elastic transformations.

• Split: 70% training, 15% validation, 15% testing, with SMOTE for class balancing.

Methodology:

• Model: ResNet50 with:

  • 50-layer residual blocks

  • ReLU activation, batch normalization

  • Global average pooling, dropout (0.5), and softmax output

• Training Setup:

  • Optimizer: Adam

  • Loss: Categorical cross-entropy

  • Training: 25 epochs, batch size 32, NVIDIA RTX 3050 GPU

  • Callbacks: Early stopping, model checkpointing, TensorBoard

Experimental Results:

• High classification accuracy across all disease categories.

• Grad-CAM used for visualizing model attention on image regions, showing accurate focus on lesion areas.

• Training/validation trends showed stable convergence without overfitting, aided by regularization and augmentation.

Conclusion

This study introduces a robust deep learning system capable of multi-class skin disease classification using CNNs. By applying a ResNet50 architecture trained on a diverse and well-labeled dataset, we achieved high accuracy and strong generalization across disease categories. The model demonstrated potential for integration into real-time diagnostic platforms, especially in resource-limited healthcare settings. In future, we aim to: • Expand the dataset with more skin tones and rare conditions. • Deploy the model in a mobile application using TensorFlow Lite. • Collaborate with dermatologists for real-world clinical validation. • Incorporate hybrid models combining CNNs with attention mechanisms. • Implement model compression techniques for edge deployment.

References

[1] Esteva, A. et al. (2017). Dermatologist-level classification of skin cancer with deep neural networks. Nature. [2] Codella, N. et al. (2019). Skin lesion analysis toward melanoma detection. IEEE J. of Biomedical and Health Informatics. [3] Gururaj, H.L. et al. (2023). DeepSkin: A Deep Learning Approach for Skin Cancer Classification. IEEE Access. [4] .Author et al. (2023). Region-of-Interest-Based Transfer Learning for Skin Disease Classification. [5] He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep Residual Learning for Image Recognition. CVPR. [6] Selvaraju, R.R., et al. (2017). Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization. ICCV. [7] Shorten, C. & Khoshgoftaar, T.M. (2019). A survey on image data augmentation for deep learning. Journal of Big Data. [8] Ronneberger, O., et al. (2015). U-Net: Convolutional Networks for Biomedical Image Segmentation. MICCAI.

Copyright

Copyright © 2025 Pratik Nagargoje, Onkar Mhamane, Anushka Kale, Abhijit Patil. 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.