Towards accurate diagnostics: deep-transfer learning in skin disease detection

Abstract

Automated skin disease diagnosis remains a significant challenge due to inherent variations in skin tone, texture, lighting, and occlusions such as hair. Addressing these diagnostic difficulties is vital for developing reliable clinical decision-support systems. This research presents an automated classification framework utilizing deep learning, transfer learning, and explainable artificial intelligence (XAI). The study utilizes the ISIC 2019 dataset, comprising images across eight diagnostic categories, including Melanoma, Nevi, and Keratosis. Six state-of-the-art convolutional neural network (CNN) architectures—Xception, DenseNet169, EfficientNetV2S, MobileNetV2, InceptionResNetV2, and ResNet101V2—were fine-tuned and evaluated. The workflow integrates advanced preprocessing, comprehensive model training, and XAI-based validation to ensure transparency in diagnostic predictions. Quantitative evaluation reveals that Xception and EfficientNetV2S outperformed other architectures, achieving near-optimal performance, with Xception reaching an area under the curve (AUC) of 100% and superior scores across all primary metrics. Conversely, MobileNetV2 demonstrated the lowest comparative efficacy, with an accuracy of 62.66% and an AUC of 78.10%. The integration of high performing CNNs with XAI techniques provides an accurate, interpretable, and scalable solution for dermatological diagnostics. This approach not only improves classification precision but also enhances clinical trust by providing visual explanations for model decisions.