The implementation of Hypernet-Ensemble Learning of Segmentation Probability for Medical Image Segmentation with Ambiguous Labels (https://arxiv.org/abs/2112.06693).

We aim to improve the segmentation probability estimation of DL networks for medical image segmentation that often suffers from the ambiguity in human annotations and anatomical/pathological structures using the hypernetwork ensemble strategy with the varying Tversky loss.

The overall hypernetwork architecture shown above is implemented in networks/hyper_resunet.py.

The hyperconvolution blocks are implemented in blocks/hyper_convolution.py.

The codes are implemented with the MONAI framework (https://monai.io/), PyTorch (https://pytorch.org/), and PyTorch Lightning (https://www.pytorchlightning.ai/).

pip install -r requirements.txt

Example Dependencies

• monai=0.7.0
• torch=1.9.1
• nibabel=3.2.1

The hypernetwork and optimizer is wrapped with PyTorch Lightning (lightning_modules/module_hyper_resunet.py).

Please see train.py for setting up the network parameters and training configurations.

Please see predict.py for inference.

The network parameters need to be the same with a trained network.

@misc{hong2021hypernetensemble,
      title={Hypernet-Ensemble Learning of Segmentation Probability for Medical Image Segmentation with Ambiguous Labels}, 
      author={Sungmin Hong and Anna K. Bonkhoff and Andrew Hoopes and Martin Bretzner and Markus D. Schirmer and Anne-Katrin Giese and Adrian V. Dalca and Polina Golland and Natalia S. Rost},
      year={2021},
      eprint={2112.06693},
      archivePrefix={arXiv},
      primaryClass={eess.IV}
} 

Please contact Sungmin Hong (MGH, HMS shong20@mgh.harvard.edu) if you have questions on the codes or the paper.

Update March 2022: I joined AWS ML Solution Lab. My new email address is hsungmin@amazon.com.