DermoCC-GAN: A new approach for standardizing dermatological images using generative adversarial networks

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Dermatological image quality is highly dependent on illumination, impacting diagnosis.
• Existing color constancy algorithms have limitations due to image assumptions.
• Variability in lighting affects dermatologist accuracy and computer-aided diagnosis systems.

Purpose of the Study:

• To propose a novel Dermatological Color Constancy Generative Adversarial Network (DermoCC-GAN).
• To overcome limitations of current color constancy algorithms in dermatological imaging.
• To formulate color constancy as an image-to-image translation problem.

Main Methods:

• Trained a generative adversarial network (GAN) with a custom heuristic algorithm.
• The GAN learned domain transfer from original to color-standardized images.
• Applied color constancy to test images with diverse illumination conditions.

Main Results:

• DermoCC-GAN outperformed state-of-the-art algorithms in normalized median intensity.
• Color-normalized images improved deep learning lesion classification (79.2% accuracy) and segmentation (90.9% dice score).
• Validated on two external datasets with highly satisfactory results.

Conclusions:

• A GAN can generalize heuristic methods for dermatological image color constancy.
• The DermoCC-GAN approach enhances dermatological image analysis.
• The method is extensible to other image analysis applications.