FedGAN: Federated diabetic retinopathy image generation.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Science

Background:

• Centralized training of deep learning models for medical diagnostics raises significant patient data privacy concerns.
• Existing methods struggle to balance the need for large datasets with strict privacy regulations like HIPAA and GDPR.

Purpose of the Study:

• To introduce FedGAN, a novel federated learning framework for generating synthetic medical images.
• To address data scarcity and privacy challenges in medical AI through secure cross-institutional collaboration.

Main Methods:

• Utilizing Generative Adversarial Networks (GANs) combined with cross-silo federated learning.
• Pretraining a Deep Convolutional GAN (DCGAN) on CT scans and fine-tuning on diabetic retinopathy datasets.
• Employing the Federated Averaging (FedAvg) algorithm to collaboratively train the GAN's components across clinical sites.

Main Results:

• FedGAN successfully generates high-quality synthetic retinal images.
• The generated images achieved a realism score of 0.43, as evaluated by a centralized discriminator.
• The framework ensures compliance with data privacy regulations (HIPAA, GDPR).

Conclusions:

• FedGAN offers a viable solution for privacy-preserving synthetic medical image generation.
• This framework facilitates secure data collaboration among institutions, overcoming data scarcity issues.
• The study demonstrates the potential of federated learning in advancing medical AI applications.