An Improved YOLOv8-Based Object Detection Algorithm for Skin Diseases

Accurate and effective diagnosis of skin diseases is crucial for clinical decision-making. However, there are still some challenges, including irregular lesion morphologies, class imbalance between rare and common types, and performance degradation in complex scenarios characterized by noise or occlusion. To address these issues, an improved skin disease detection algorithm is proposed based on YOLOv8n, which features three core innovations. First, it integrates deformable large kernel attention (D-LKA) into deep backbone layers to capture global contextual relationships of lesions; then it embeds DCNv3 deformable convolutions in mid-layers to adaptively sample irregular lesion boundaries; finally, it designs an EMA-Slide Loss function to dynamically weight hard-to-classify samples, thereby reducing bias toward common categories. After evaluating on the International Skin Imaging Collaboration (ISIC) dataset (with labels validated by board-certified dermatologists), the algorithm can achieve 96.58% mAP50 and 88.32% mAP50-95, 2.44% and 2.52% higher than the baseline YOLOv8n, respectively. It maintains a real-time inference speed of 31 ms per image, making it suitable for edge devices such as portable dermatoscopes. Supporting nine common types of skin diseases, with extensibility to accommodate rare types and multi-modal data fusion, this work provides a clinically actionable tool for automated skin lesion analysis, bridging the gap between algorithmic performance and real-world clinical demands.