meddlr#

Meddlr is a config-driven ML framework built to simplify medical image reconstruction and analysis problems.

⚡ QuickStart#

# Install Meddlr with basic dependencies
pip install meddlr

# Install Meddlr with all dependencies (e.g. pretrained models, benchmarking)
pip install 'meddlr[all]'

Installing locally: For local development, fork and clone the repo and run pip install -e ".[alldev]"

Installing from main: For most up-to-date code without a local install, run pip install "meddlr @ git+https://github.com/ad12/meddlr@main"

Configure your paths and get going!

import meddlr as mr
import os

# (Optional) Configure and save machine/cluster preferences.
# This only has to be done once and will persist across sessions.
cluster = mr.Cluster()
cluster.set(results_dir="/path/to/save/results", data_dir="/path/to/datasets")
cluster.save()
# OR set these as environment variables.
os.environ["MEDDLR_RESULTS_DIR"] = "/path/to/save/results"
os.environ["MEDDLR_DATASETS_DIR"] = "/path/to/datasets"

Detailed instructions are available in Getting Started.

Visualizations#

Use MeddlrViz to visualize your medical imaging datasets, ML models, and more!

pip install meddlr-viz

A gallery of images from the BRATS dataset

🐘 Model Zoo#

Easily serve and download pretrained models from the model zoo. A (evolving) list of pre-trained models can be found here, on HuggingFace 🤗, and in project folders.

To use them, pass the URLs for the config and weights (model) files to mr.get_model_from_zoo:

import meddlr as mr

model = mr.get_model_from_zoo(
  cfg_or_file="https://huggingface.co/arjundd/vortex-release/resolve/main/mridata_knee_3dfse/Supervised/config.yaml",
  weights_path="https://huggingface.co/arjundd/vortex-release/resolve/main/mridata_knee_3dfse/Supervised/model.ckpt",
)

📓 Projects#

Check out some projects built with meddlr!

✏️ Contributing#

Want to add new features, fix a bug, or add your project? We’d love to include them! See CONTRIBUTING.md for more information.

Acknowledgements#

Meddlr’s design for rapid experimentation and benchmarking is inspired by detectron2.

About#

If you use Meddlr for your work, please consider citing the following work:

@article{desai2021noise2recon,
  title={Noise2Recon: A Semi-Supervised Framework for Joint MRI Reconstruction and Denoising},
  author={Desai, Arjun D and Ozturkler, Batu M and Sandino, Christopher M and Vasanawala, Shreyas and Hargreaves, Brian A and Re, Christopher M and Pauly, John M and Chaudhari, Akshay S},
  journal={arXiv preprint arXiv:2110.00075},
  year={2021}
}