GaNDLF-Synth

The Generally Nuanced Deep Learning Framework - Synthesis (GaNDLF-Synth) for reproducible and automated deep generative modeling in medical imaging.

Why use GaNDLF-Synth?

GaNDLF-Synth was developed to lower the barrier to AI, enabling reproducibility, translation, and deployment regarding usage of generative models in medical imaging. It is an extension of the GaNDLF framework, which is a part of the MLCommons initiative. GaNDLF-Synth aims to extend the capabilities of GaNDLF to include generative models, such as GANs, VAEs, and diffusion models, while adhering to the same principles.
As an out-of-the-box solution, GaNDLF alleviates the need to build from scratch. Users may kickstart their project by modifying only a configuration (config) file that provides guidelines for the envisioned pipeline and CSV inputs that describe the training data.

Range of GaNDLF-Synth functionalities:

• Supports multiple
  • Deep Generative model architectures
  • Channels/modalities
  • Labeling schemes (per patient, per custom class, unlabeled)
• Support of multiple loss, optimizers, scheduler, data augmentation, and evaluation metrics via interfacing GaNDLF
• Multi-GPU and multi-node training and inference support, integrating DistributedDataParallel DDP and deepspeed; parallelism present both on the model and data level
• Leverages robust open source software - Pytorch Lightning, monai-generative
• Zero-code needed to train robust models and generate synthetic data
• Low-code requirement for customization and addition of custom models and training logic
• Automatic mixed precision support

Table of Contents

• Getting Started
• Application Setup
• Usage
  • Customize the training and inference
• Extending GaNDLF
• FAQ
• Acknowledgements

Citation

Please cite the following article for GaNDLF-Synth:

@misc{pati2024gandlfsynthframeworkdemocratizegenerative,
      title={GaNDLF-Synth: A Framework to Democratize Generative AI for (Bio)Medical Imaging}, 
      author={Sarthak Pati and Szymon Mazurek and Spyridon Bakas},
      year={2024},
      eprint={2410.00173},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2410.00173}, 
}

Contact

GaNDLF developers can be reached via the following ways: - GitHub Discussions - Email