Primary Classification of Skin Diseases Using an Explainable Multimodal VLM with a DBSCAN-Centroid-Based Confidence Score

This paper presents an explainable multimodal vision–language framework for the primary classification of skin diseases. Using compact vision–language models (VLMs)—Gemma 3 4B and Qwen 2.5 VL 7B—the system integrates synthetic skin tumor and lesion images with natural-language disease descriptions, grounding its predictions in lay-accessible dermatologic concepts to improve interpretability. Low-rank adaptation (LoRA) fine-tuning on the AI-Hub synthetic skin-tumor dataset demonstrates the feasibility of deploying such models in resource-constrained environments. Model performance is evaluated using quantitative metrics, including accuracy, precision, recall, and F1-score, and a DBSCAN-centroid-based semantic confidence-scoring method is introduced to estimate cluster similarities in the image-embedding space. The experimental results show that lightweight multimodal VLMs can achieve stable and accurate performance on primary skin disease classification, indicating their potential as explainable, AI-assisted tools for dermatologic decision support.