Multi-Class Skin Disease Classification using Transfer Learning and Explainable AI

Abstract

This study investigates the application of transfer learning and Explainable Artificial Intelligence (XAI) for multi-class skin disease classification. The dataset was obtained from the Kaggle Skin Diseases Image Dataset and consists of 29,153 original images spanning 10 skin disease classes. To reduce the bias introduced by class imbalance, the dataset was balanced through directed undersampling, resulting in 12,000 images, with 1,200 images per class. Three pretrained convolutional neural network (CNN) architectures—EfficientNetB0, ResNet50, and DenseNet201—were implemented and evaluated using a confusion matrix, accuracy, precision, recall, and F1-score. The experimental results demonstrate that DenseNet201 achieved the highest classification performance, with an accuracy of 0.8779, precision of 0.8751, recall of 0.8748, and F1-score of 0.8745, outperforming ResNet50 (accuracy: 0.8629) and EfficientNetB0 (accuracy: 0.8269). Model interpretability was investigated using Grad-CAM, SHAP, and LIME. Grad-CAM highlighted that the models primarily focused on the central and peripheral regions of skin lesions during prediction. SHAP identified the dominant contribution of lesion regions and pigmentation patterns to the classification process, while LIME emphasized the importance of local superpixels associated with lesion boundaries, color, and texture in supporting the model's predictions. The findings indicate that combining transfer learning with Explainable AI provides a promising foundation for developing clinical decision support systems for dermatological image classification. Future research should incorporate external dataset validation, more robust class balancing strategies, and clinical interpretation by dermatology experts to facilitate the deployment of such systems in real-world healthcare settings.