imedigan

imedigan provides access to pretrained generative models for medical image synthesis to support clinical decision support systems and medical imaging research.

Key Features:

• Pretrained Generative Models: A repository of 21 pretrained models based on nine distinct generative adversarial network (GAN) architectures trained on 11 datasets spanning mammography, endoscopy, x‑ray, and MRI.
• Modular Components: Modular components for execution, visualization, search and ranking, and contribution enable running models, visualizing synthetic outputs, finding and ranking models, and adding new models.
• Framework-agnostic Architecture: A framework-agnostic architecture that supports integration with different deep learning frameworks.
• Evaluation and Metrics Support: Support for extraction and analysis of Fréchet Inception Distance (FID) with consideration of image normalization and radiology-specific feature extractors.
• Scalability: Design intended to scale to additional models and datasets.

Scientific Applications:

• Community-wide Data Sharing: Enable sharing of restricted data via synthetic images to facilitate collaboration across the medical research community.
• Evaluation Metrics Analysis: Facilitate investigation of generative model evaluation metrics such as FID and assessment of variability due to image normalization and radiology-specific feature extractors.
• Clinical Task Improvement: Provide synthetic data to augment training sets for clinical downstream machine learning tasks.

Methodology:

Access and reuse pretrained GANs to generate synthetic medical images for data augmentation and domain adaptation, and compute Fréchet Inception Distance (FID) using different image normalization schemes and radiology-specific feature extractors.