Controllable synthesis of dermoscopic images using diffusion models for enhanced computer aided diagnosis and...

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Limited and imbalanced dermoscopic datasets pose challenges for Computer Aided Diagnosis/Detection (CAD) systems in skin lesion analysis.
• Advanced data augmentation techniques are crucial for improving CAD model performance.

Purpose of the Study:

• To introduce DiDGen, an innovative method for generating high-quality dermoscopic images using text-to-image Diffusion models.
• To enhance the performance of CAD systems for skin lesion analysis through improved data augmentation.

Main Methods:

• Proposed DiDGen, a method utilizing text-to-image Diffusion models for dermoscopic image generation.
• Introduced DermPrompt, a dynamic prompting framework leveraging large language models for attribute-rich text prompts.
• Implemented a novel region-aware fine-tuning approach and a training-free pipeline for synthesizing lesion-mask pairs.

Main Results:

• DiDGen demonstrated superior image fidelity and diversity compared to existing generative methods.
• Downstream classifiers and segmentation models showed average improvements of 2.32% in F1 score and 3.16% in IoU score.
• These improvements were achieved with a single fine-tuning process.

Conclusions:

• DiDGen offers an efficient solution for augmenting dermoscopic datasets.
• The proposed method effectively advances skin lesion diagnosis by enhancing CAD system performance.