Diagnosing and Typing Scalp Conditions with Transformer-Based Model

Scalp health issues are increasingly prevalent due to stress, environmental pollution, and dietary changes, but current diagnostic methods rely on subjective evaluations, limiting early detection and personalized treatment development. This study proposes a multi-class classification framework for scalp condition analysis using over 100,000 publicly available scalp images. We applied image preprocessing techniques including Gaussian filtering and bottom-hat transformation, along with data augmentation to address class imbalance. The Vision Transformer (ViT-B/16) architecture was employed for classifying six major scalp diseases and four severity levels. Additionally, K-means clustering was used for unsupervised scalp condition stratification to enhance diagnostic interpretability. The proposed preprocessing and augmentation techniques improved ViT-B/16 classification accuracy from 74.1% to 85.6%. K-means clustering revealed three distinct scalp types: a sebaceous cluster characterized by excess sebum and interfollicular erythema, a keratinous cluster defined by fine keratin predominance, and a complex cluster exhibiting mixed symptoms. These classification and clustering results provide a practical foundation for objective scalp condition categorization and support the development of standardized diagnostic protocols, offering potential for more accurate and personalized scalp health assessments.