An Improved YOLOv8-Based Object Detection Algorithm for Skin Diseases

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Accurate skin disease diagnosis is vital but challenged by irregular lesion morphology, class imbalance, and complex imaging conditions.
• Existing algorithms often struggle with these complexities, impacting clinical decision-making.

Purpose of the Study:

• To develop an improved skin disease detection algorithm addressing current diagnostic challenges.
• To enhance the accuracy and robustness of automated skin lesion analysis.

Main Methods:

• An improved skin disease detection algorithm based on YOLOv8n was developed.
• Key innovations include deformable large kernel attention (D-LKA), DCNv3 deformable convolutions, and an EMA-Slide Loss function.
• The algorithm was evaluated on the International Skin Imaging Collaboration (ISIC) dataset.

Main Results:

• The algorithm achieved 96.58% mAP50 and 88.32% mAP50-95 on the ISIC dataset, outperforming the baseline YOLOv8n.
• Demonstrated a real-time inference speed of 31 ms per image, suitable for edge devices.
• The method effectively handles irregular lesion morphologies and class imbalance.

Conclusions:

• The proposed algorithm offers a clinically actionable tool for automated skin lesion analysis.
• It significantly improves diagnostic accuracy and robustness compared to the standard YOLOv8n.
• The algorithm is suitable for real-time applications and extensible for rare disease types and multi-modal data.