Matwo-CapsNet

Implementation of the MICCAI 2019 Paper "Matwo-CapsNet: a Multi-Label Semantic Segmentation Capsules Network".

Matwo-Caps

SRC Multi-Label segmentation Dice score

Method	Parameters	Loss	Left lung	Right lung	Left clavicule	Right clavicule	Heart
U-Net	42K	Softmax	97.36	97.87	90.87	90.64	94.49
SegCaps	2,129K	weighted Softmax	21.18	35.79	4.49	2.93	32.83
SegCaps	2,129K	weighted Spread loss	30.74	0	0.06	0	23.23
MatVec-CapsNet O_r	43K	Spread loss	95.57	96.43	82.89	82.56	92.37
MatVec-CapsNet D_r	43K	Spread loss	96.60	97.15	86.41	86.38	93.42
Matwo-CapsNet	43K	Spread loss	97.01	97.45	88.32	87.82	94.37

Architecture

Dataset preprocessing

Download the files from :

Extract the archives in ./Dataset before running ./Dataset/dataset_preprocessing.py

Train models

Run main_train_and_test.py to train the network. You can run any of the paper experiment by simply uncommenting the different settings at the end of the script.

Train and test other datasets

In order to train and test on other datasets, modify the dataset.py file. See the example files and documentation for the specific file formats. Set the parameter save_debug_images = True in order to see, if the network input images are reasonable.

Citation

If you use this code for your research, please cite our Paper:

@inproceedings{Bonheur2019,
  title     = {Matwo-CapsNet: a Multi-Label Semantic Segmentation Capsules Network},
  author    = {Bonheur, Savinien and {\v{S}}tern, Darko and Payer, Christian and Pienn, Michael and Olschewski, Horst and Urschler, Martin},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  doi       = {978-3-030-32254-0_74},
  pages     = {664-672},
  year      = {2019},
}

References

This code make use of the following projects:

Framework

@inproceedings{Payer2018a,
  title     = {Multi-label Whole Heart Segmentation Using {CNNs} and Anatomical Label Configurations},
  author    = {Payer, Christian and {\v{S}}tern, Darko and Bischof, Horst and Urschler, Martin},
  booktitle = {Statistical Atlases and Computational Models of the Heart. ACDC and MMWHS Challenges. STACOM 2017},
  doi       = {10.1007/978-3-319-75541-0_20},
  pages     = {190--198},
  year      = {2018},
}

Segcaps

@article{lalonde2018capsules,
  title={Capsules for Object Segmentation},
  author={LaLonde, Rodney and Bagci, Ulas},
  journal={arXiv preprint arXiv:1804.04241},
  year={2018}
}

Special Requirement:

-cachetools 2.1.0

savinienb / Matwo-CapsNet

readme