Optimizing Dice score and Jaccard Index for Medical Image Segmentation

This repository open sources the code and models belonging to the public datasets used in:

• Optimizing the Dice Score and Jaccard Index for Medical Image Segmentation: Theory and Practice

   Bertels, J., Eelbode, T., Berman, M., Vandermeulen, D., Maes, F., Bisschops, R., & Blaschko, M. B. (2019, October).
   Optimizing the Dice score and Jaccard index for medical image segmentation: Theory and practice.
   In International Conference on Medical Image Computing and Computer-Assisted Intervention (pp. 92-100). Springer, Cham.
  

  Links: Springer | arXiv

• Optimization for Medical Image Segmentation: Theory and Practice when evaluating with Dice Score or Jaccard Index

   Eelbode, T., Bertels, J., Berman, M., Vandermeulen, D., Maes, F., Bisschops, R., & Blaschko, M. B. (2020).
   Optimization for medical image segmentation: Theory and practice when evaluating with dice score or jaccard index.
   IEEE Transactions on Medical Imaging.
  

  Links: IEEE

If you find the code useful for your research, please consider citing these works.

Get started

1. To obtain or get more information about the original data please go to:

   • BRATS 2018: https://www.med.upenn.edu/sbia/brats2018/data.html
   • ISLES 2017: http://www.isles-challenge.org/ISLES2017/
   • ISLES 2018: http://www.isles-challenge.org
   • WMH 2017: https://wmh.isi.uu.nl/data/

2. The original public data was resampled to a 2 mm isotropic voxel size (using scipy.ndimage.zoom with order=1).

3. Make Keras model using ./unet_generalized.py and using the information in .//model_info.txt (models are the same in general, but the model for ISLES_2018 has less parameters).

4. To do predictions the information in .//other_info.txt is necessary. For example for BRATS 2018:

   • Order of inputs (in feature dimension): ["FLAIR", "T1", "T1_CE", "T2"]
   • Normalize inputs using: (input - shift) / scale shifts: [420.884688397679, 568.7868683246469, 639.4882077323609, 629.919352934067] scales: [1320.6450427506038, 1160.6822019612432, 1181.144425870453, 1363.6117673325714]
   • Extract central patch of spatial size: [136, 136, 82]
   • Masking the predictions with FLAIR > 0 (can be done by setting mask argument to True when creating the model (see step 3.) and providing this as an extra input)

5. Validation splits can be found in .//validation_splits.txt

Contacts

• Bertels Jeroen | GitHub: JeroenBertels | ResearchGate: Jeroen Bertels
• Eelbode Tom | GitHub: TomEelbode | ResearchGate: Tom Eelbode | LinkedIn: tomeelbode

License

All the code in this repository is covered by the LICENSE. Please refer to the LICENSE for details.