Diffusion-based skin disease data augmentation with fine-grained detail preservation and interpolation for data...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Dermatology

Background:

• Data scarcity is a significant challenge in training AI models for skin disease diagnosis.
• Existing generative models may lose critical details in synthetic medical images.

Purpose of the Study:

• To develop an advanced data augmentation technique for skin disease diagnosis research.
• To enhance the quality and diversity of synthetic medical images using diffusion models.

Main Methods:

• Enhanced a Stable Diffusion (Latent Diffusion Model) by improving encoder/decoder structures and incorporating lesion masks.
• Utilized multi-level embeddings with a CLIP encoder for detailed image representation.
• Employed pre-trained segmentation and inpainting models for generating normal skin regions and interpolating visual characteristics.

Main Results:

• Generated high-quality synthetic skin images with improved detail preservation.
• Demonstrated enhanced classification performance on seven skin diseases when combining synthetic and real data.
• Validated the effectiveness of the diffusion-based augmentation technique in addressing data scarcity.

Conclusions:

• The proposed diffusion-based data augmentation method effectively increases data diversity for skin disease research.
• The technique shows potential for improving diagnostic accuracy in medical AI applications.
• Synthetic data generated can serve as valuable reference material, despite current clinical limitations.