Key points are not available for this paper at this time.
Skin diseases are some of the most common medical conditions worldwide, ranging from mild disorders to life-threatening cancers. Their accurate diagnosis remains challenging because of high visual variability and overlapping clinical symptoms. This systematic review, covering studies published between 2018 and 2025, evaluates computer vision, machine learning (ML), deep learning (DL), transfer learning, and hybrid approaches for dermatological disease classification. It specifically focuses on image-based (single-modal) artificial intelligence systems for dermatological diagnosis. Studies that primarily relied on textual clinical notes, electronic health records, or multimodal fusion of imaging and non-imaging data were outside the scope of this review. A total of 84 peer-reviewed studies were analyzed, with focus on datasets, model architectures, validation techniques, and reported outcomes. The review followed PRISMA guidelines and involved systematic searches across several major scientific databases. Studies were screened, assessed for quality, and systematically synthesized. The findings highlight the prevalence of convolutional neural networks (CNNs). Lightweight architectures, including MobileNet and EfficientNet variants, are frequently reported for achieving competitive accuracy with lower computational complexity, supporting their use in resource-constrained clinical settings. Transfer learning and hybrid approaches further improve adaptability and robustness across diverse datasets. Despite promising results, challenges remain in dataset diversity, fairness across skin tones, and limited clinical validation. These findings underscore the importance of responsible and deployment-aware AI design to support accurate, efficient, and equitable dermatological decision-support systems.
Aljohani et al. (Tue,) studied this question.