Deep Transfer Learning with Ensemble Models for Skin Disease Classification: A Case Study on Melanoma and Eczema

Accurate and early classification o f s kin l esions is essential for the timely diagnosis and treatment of dermatological diseases such as eczema and melanoma. However, conventional automated methods often face challenges due to high visual similarity between lesions and the limited diversity and size of available datasets. To address these issues, this study proposes an ensemble deep learning framework that leverages transfer learning to improve binary skin lesion classification performance. The proposed approach integrates three pre-trained convolutional neural network (CNN) architectures—MobileNetV2, ResNet50, and VGG16— each fine-tuned on a curated dataset of eczema and melanoma images. A soft voting ensemble strategy is employed to aggregate predictions from the individual models, aiming to enhance classification robustness and mitigate the weaknesses of any single model. Experimental results demonstrate that the ensemble model achieves a classification accuracy of 95.43%, a macro F1-score of 0.95, and an AUC of 1.00, outperforming ResNet50 and VGG16, and matching the superior performance of MobileNetV2.