Exploring Web Images to Enhance Skin Disease Analysis Under A Computer Vision Framework.

Area of Science:

• Dermatology
• Computer Vision
• Artificial Intelligence

Background:

• Accurate skin disease diagnosis is crucial for effective treatment.
• Collecting sufficient labeled dermatological images for training AI models is challenging.
• Existing methods often focus on limited disease types and rely on small, private datasets.

Purpose of the Study:

• To design an automatic and effective visual analysis framework for skin disease recognition.
• To address the challenge of insufficient labeled data by leveraging external web image data.
• To develop a novel transfer learning model capable of incorporating knowledge from diverse image sources.

Main Methods:

• Constructed a target domain using professional dermatological website images and a source domain using large-scale web images.
• Developed a transfer learning model that integrates knowledge from both domains.
• Bridged the domain distribution gap using a linear combination of Gaussian kernels.
• Employed deep learning models for high-level image representation learning, moving beyond low-level features.

Main Results:

• The proposed transfer learning model demonstrated superior performance compared to state-of-the-art methods.
• The framework successfully recognized thousands of common skin diseases using publicly accessible web images.
• Experiments on a real-world dataset validated the effectiveness and robustness of the approach.

Conclusions:

• The novel transfer learning approach offers an effective solution for skin disease visual analysis, overcoming data limitations.
• The method is repeatable and extendable to other disease types, promoting wider research and application.
• This work advances automated dermatological diagnostics by utilizing readily available web data.