Deep Transfer Learning–based Multi-Class Skin Lesion Recognition and Segmentation for Clinical Decision Support

Accurate and timely identification of skin lesions is critical for improving clinical outcomes in dermatological practice, particularly for conditions that exhibit subtle visual differences. This study proposes a deep transfer learning–based framework for automated multi-class skin lesion recognition and instance segmentation using the YOLOv12-seg architecture. A diverse and balanced dataset comprising 2,039 dermoscopic and clinical images across four lesion categories— Melanocytic Nevi, Vascular Lesions, Cowpox, and Dermatofibroma—was curated from multiple publicly available sources and annotated for supervised learning. The model was trained using an 87%–8%–4% train–validation–test split and evaluated using standard detection and segmentation metrics. On the held-out test set, the proposed framework achieved an overall mAP@50 of 0.816 and an mAP@50–95 of 0.514, with corresponding precision and recall values of 0.778 and 0.750, respectively. Class-wise analysis demonstrated particularly strong performance for Vascular Lesions and Melanocytic Nevi, while qualitative results confirmed accurate lesion localization and well-defined segmentation boundaries across diverse imaging conditions. The model exhibited high computational efficiency, achieving millisecond-level inference, and was successfully deployed via a Streamlit-based web application for real-time visualization and interaction. These results demonstrate the effectiveness, robustness, and practical applicability of the proposed YOLOv12-seg framework as a reliable clinical decision support tool for automated skin lesion analysis.