Beyond a single mode: GAN ensembles for diverse medical data generation

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Generative AI in medical imaging faces challenges in generating high-fidelity and diverse synthetic data.
• Generative Adversarial Networks (GANs) are promising but suffer from mode collapse and poor data distribution coverage.
• This study explores GAN ensembles to overcome these limitations and enhance synthetic medical image quality.

Purpose of the Study:

• To investigate the use of GAN ensembles for improved synthetic medical image generation.
• To address the trilemma of fidelity, diversity, and efficiency in generative AI for medical imaging.
• To enhance the quality and utility of synthetic medical images for clinical and research applications.

Main Methods:

• Formulated a multi-objective optimization problem for selecting GAN ensembles balancing fidelity and diversity.
• Ensured ensemble models contribute uniquely to the synthetic data space, minimizing redundancy.
• Evaluated 22 GAN architectures across three medical imaging datasets, using 110 unique configurations.

Main Results:

• Selected GAN ensembles generated synthetic medical images with improved fidelity and diversity, closely matching real data distributions.
• Downstream models trained on synthetic data achieved comparable or slightly higher accuracy than those trained on real data alone.
• Synthetic images served as effective data augmentation, enhancing class balance and diversity.

Conclusions:

• GAN ensembles provide a robust solution to the fidelity-diversity-efficiency trade-off in medical image synthesis.
• Integrating complementary GAN models improves the representativeness and utility of synthetic medical data.
• This approach has the potential to advance diagnostic AI applications in healthcare.