medigan: a Python library of pretrained generative models for medical image synthesis.

Area of Science:

• Medical Artificial Intelligence
• Machine Learning
• Medical Imaging

Background:

• Deep learning models in clinical decision support require substantial data.
• Synthetic data generation using generative models can augment limited datasets.
• Barriers to generative model adoption include data scarcity and training complexity.

Purpose of the Study:

• To reduce entry barriers for researchers and clinicians in utilizing synthetic data.
• To explore generative model sharing as a solution for data accessibility.
• To enable broader access to and benefit from synthetic data generation.

Main Methods:

• Developed medigan, an open-source, framework-agnostic Python library for pretrained generative models.
• Designed medigan with modular components for model execution, visualization, search, ranking, and contribution.
• Integrated pretrained models across various medical imaging modalities (mammography, endoscopy, X-ray, MRI).

Main Results:

• Integrated 21 pretrained generative models using nine GAN architectures trained on 11 datasets.
• Demonstrated medigan's scalability and usability through diverse applications.
• Analyzed generative model evaluation metrics, highlighting variability in Fréchet Inception Distance (FID) based on normalization and feature extractors.

Conclusions:

• Medigan empowers researchers to easily create, augment, and adapt training data using synthetic data.
• The library accelerates clinical machine learning model development and validates generative model sharing.
• Established standards for assessing and reporting metrics like FID in synthetic image generation.