ARTIFICIAL INTELLIGENCE FOR SKIN DISEASE DETECTION: A REVIEW OF MACHINE LEARNING AND DEEP LEARNING MODELS, PUBLIC DATASETS, PERFORMANCE TRENDS, AND CLINICAL CHALLENGES

Abstract

Skin diseases are among the most common health conditions globally, affecting individuals across all age groups and demographics. Accurate and timely diagnosis is essential to prevent complications and improve patient outcomes. However, dermatological expertise is not always readily available, particularly in remote or resource-limited areas and manual diagnosis by dermatologists can be time-consuming and subjective. In this context, machine learning (ML) and deep learning (DL) techniques have emerged as promising tools for automated skin disease detection and classification, offering the potential to enhance diagnostic accuracy and accessibility. This review paper explores various machine and deep learning-based approaches for skin disease classification, focusing on publicly available datasets, model architectures, and performance metrics. Several datasets, such as ISIC (International Skin Imaging Collaboration) [1], HAM10000 (Human Against Machine with 10,000 training images) [2], and DermNet [3], have been widely used in research. Studies leveraging Convolutional Neural Networks (CNNs), including ResNet, DenseNet, and EfficientNet, have achieved diagnostic accuracies exceeding 90% in some cases. For instance, Esteva et al. (2017) demonstrated that a deep neural network could classify skin cancer with dermatologist-level accuracy [4]. Similarly, transfer learning and data augmentation techniques have been employed to address dataset limitations, such as class imbalance and insufficient samples. A systematic search strategy was employed to identify peer-reviewed studies published between January 2015 and April 2025 from multiple databases, including PubMed, IEEE Xplore, Scopus, Web of Science, and Google Scholar. Studies were evaluated based on algorithm type, dataset usage, performance metrics, and clinical applicability. Commonly used public datasets such as ISIC and HAM10000 were found to play a crucial role in training and validating these systems. The review categorizes the selected literature into ML-based and DL-based models, highlighting key trends, strengths, and limitations of each approach. Additionally, it discusses challenges such as data imbalance, model generalizability, and interpretability. The findings suggest that while deep learning models—especially convolutional neural networks, offer superior performance in image classification tasks, the integration of explainable AI and diverse datasets remains essential for clinical translation. This paper aims to guide future research by identifying gaps and proposing directions for the development of more robust, interpretable, and clinically deployable skin disease detection systems