GestaltGAN

GestaltGAN generates photorealistic portraits of individuals with rare genetic disorders using a modified StyleGAN architecture, addressing privacy concerns in medical imaging and providing synthetic data for publications, education, and AI training.

Key Features:

• Condition-Specific Portrait Generation: GestaltGAN produces random condition-specific portraits for multiple rare genetic disorders using a modified StyleGAN architecture.
• Average Patient Portrait Synthesis: The technique generates a sharp and detailed average patient portrait for a given disorder.
• Training Data Enhancement: REAL-ESRGAN increases the resolution of low-quality portraits and colorizes black-and-white images from the training data.
• Data Alignment and Augmentation: Training data is aligned and cropped to achieve a uniform format, and an unaffected class is introduced to augment the model's understanding of human facial features.
• Expert Validation: Generated portraits were tested with 63 human experts to assess validity and clinical accuracy.
• Privacy Preservation: The generated portraits capture characteristic disorder features while preserving patient privacy.

Scientific Applications:

• Scientific Publications: Synthetic portraits provide visualizations for publications without exposing patient identities.
• Medical Education: Generated images serve as educational materials for teaching rare genetic disorders.
• AI Training Data Augmentation: Synthetic portraits augment training data for deep learning methods such as GestaltMatcher.

Methodology:

GestaltGAN modifies a StyleGAN architecture and trains it on the 20 most frequent disorders from the GestaltMatcher database. REAL-ESRGAN is used to increase resolution of low-quality training portraits and colorize black-and-white images. The training data is aligned and cropped to a uniform format, with an unaffected class introduced to augment facial feature understanding. A technique generates sharp, detailed average patient portraits for given disorders.