Sharmin Islam Shroddha, Rashedur Rahman, Ashraful Islam, Md Junayed Hossain, Sefatul Wasi, Saadia Binte Alam, M Ashraful Amin
2025 IEEE 23rd Student Conference on Research and Development (SCOReD)
In: 2025 IEEE 23rd Student Conference on Research and Development (SCOReD)
IEEE, pp. 1-6
Classifying skin diseases from dermoscopic images is important for the early detection and treatment of skin cancer, as timely diagnosis can significantly improve patient outcomes. In this study, we perform a comparative analysis of three widely used deep learning models EfficientNetB0, ResNet50, and Vision Transformer (ViT) for automated classification of dermoscopic images from the HAM10000 dataset. Since the dataset is highly imbalanced across the seven disease categories, we applied data augmentation and downsampling techniques to create a balanced set of 2,000 images per class. EfficientNetB0 and ResNet50 were trained using transfer learning and fine-tuning strategies, while the ViT was implemented and trained from scratch with customized hyperparameters. The models were evaluated using accuracy, precision, recall, F1-score, and AUC to provide a comprehensive assessment of their performance. Experimental results demonstrate that EfficientNetB0 achieved the highest performance with 91% accuracy and an average AUC of 0.99, closely followed by ViT with 90% accuracy and an AUC of 0.99, while ResNet50 reached 86% accuracy with an AUC of 0.98. These findings suggest that lightweight architectures such as EfficientNetB0 can outperform deeper convolutional networks and achieve results comparable to transformer-based approaches.
Assistant Professor
Department of Computer Science and Engineering
Independent University, Bangladesh (IUB)Human-Computer Interaction, AI for Social Good, AI for Public Health, AI for Impact