The Construction of a Multi-Modal Dermatology Database and Secondary Clustering Analysis of Rosacea

The primary objective of this study is to develop a comprehensive dermatological database enriched with extensive clinical information and to automatically assign pseudo-labels to dermoscopic images using deep clustering technology, thereby enhancing the efficiency of diagnostic processes and reducing the burden on medical resources. Although deep clustering effectively aggregates similar cases, the precision of the resulting categories does not yet meet clinical demands. To address this, we implemented a secondary clustering analysis using a minimal amount of precisely labeled data for weakly supervised learning, further refining disease classification. For example, the classification of acne was significantly improved by secondary clustering, which precisely distinguished rosacea from initial clustering categories. This method not only optimized disease classification outcomes but also laid a solid foundation for in-depth research and clinical applications in dermatology. Moving forward, we plan to expand the database, incorporate more types of clinical data, and explore advanced deep learning techniques to enhance the accuracy of clustering and classification, further supporting the diagnosis and treatment of skin diseases.