Description

Welcome to medigan: medical generative (adversarial) networks

Aim and Scope

medigan focuses on automating medical image dataset synthesis using GANs.

These datasets can again be used to train diagnostic or prognostic clinical models such as disease classification, detection and segmentation models.

Despite this current focus, medigan, is readily extendable to any type of modality and any type of generative model.

Note

More detail is available in the medigan article now available as preprint. The article also includes experiments with insights into FID as generative model evaluation metric.

Core Features

  • Researchers and ML-practitioners can conveniently use an existing model in medigan for synthetic data augmentation instead of having to train their own generative model each time.

  • Users can search and find a model using search terms (e.g. “Mammography, 128x128, DCGAN”) or key value pairs (e.g. key = “modality”, value = “Mammography”)

  • Users can explore the config and information (metrics, use-cases, modalities, etc) of each model in medigan

  • Users can generate samples using a model

  • Users can also get the generate_method of a model that they may want to use dynamically inside their dataloaders

  • Model contributors can share and disseminate their generative models thereby augmenting their reach.

Architecture and Workflows

Architectural overview and main workflows

Architectural overview including main workflows consisting of (a) library import and initialisation, (b) generative model search and ranking, (c) sample generation, and (d) generative model contribution.

Notes

  • Each model in medigan has its own dependencies listed in the global.json model metadata/config file. Setting install_dependencies=True, in the generate() and related methods, triggers an automatic installation of the respective model’s python dependencies (e.g. numpy. torch, etc), i.e., to the user’s active python environment.

  • Running the generate() and related methods for synthetic data generation will trigger the download of the respective generative model to the user’s local directory from where the code is run (i.e. the current CLI path).

Issues

In case you encounter problems while using medigan or would like to request additional features, please create a new issue and we will try to help.