Synthetic skin image generation using a physics-based, object-to-image computational pipeline.

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Patient datasets for skin imaging AI development often suffer from limitations such as small size and poor representativeness.
• Existing datasets may lack rare cases and comprehensive annotations crucial for robust AI model training.

Purpose of the Study:

• To provide an extended overview and evaluation of the S-SYNTH skin simulation approach.
• To assess the utility of S-SYNTH for generating synthetic skin images to aid artificial intelligence (AI) algorithm development in dermatology.

Main Methods:

• The S-SYNTH approach utilizes a knowledge-based skin object model for controlled variation in skin appearance (color, hair, lesions, blood fraction).
• The study investigates the impact of these model variations on AI models for skin lesion segmentation.

Main Results:

• Synthetic data generated by S-SYNTH exhibit similar comparative trends to patient dermatologic images.
• The approach shows potential in mitigating biases and limitations inherent in current patient datasets.
• S-SYNTH demonstrated a novel use case in enhancing diffusion model performance for dermatological applications.

Conclusions:

• Synthetic images generated via simulation can effectively address limitations of available patient datasets in skin imaging.
• The S-SYNTH approach allows for precise control over the parameter space in dataset simulation.
• This enables the creation of rare examples and annotations often absent in real-world patient datasets.