Voldemort108X/myops20

Fully Automated Deep Learning-based Segmentation of Normal, Infarcted and Edema Regions from Multiple Cardiac MRI Sequences

MyoPS 2020: Fully Automated Deep Learning-based Segmentation of Normal, Infarcted and Edema Regions from Multiple Cardiac MRI Sequences

[Paper] [Challenge website]

Overview

Bibtex

@inproceedings{zhang2020fully,
  title={Fully automated deep learning based segmentation of normal, infarcted and edema regions from multiple cardiac MRI sequences},
  author={Zhang, Xiaoran and Noga, Michelle and Punithakumar, Kumaradevan},
  booktitle={Myocardial Pathology Segmentation Combining Multi-Sequence CMR Challenge},
  pages={82--91},
  year={2020},
  organization={Springer}
}

Dataset

Please refer to challenge website [link] for dataset access. The dataset contains three folders: train25, train25_myops_gd, test20.

Environment

1. UNet:

conda env create -f myops_unet.yml

2. Mask-RCNN and UNet++:

conda env create -f myops_mrcnn_unetpp.yml

Default directory structure

├── Data                   
|   ├── Original_data       # Place the downloaded dataset here
|   |   ├── train25
|   |   ├── train25_myops_gd
|   |   ├── test20
├── mask_rcnn_coco.h5       # Downloaded pre-trained mask_rcnn weights
├── config.py
├── data_creator.py
├── ...

Setup

1. Data creator including random warping augmentation

python data_creator.py 

2. Train networks

   • Train UNet for LV_BP, RV_BP, LV_NM, LV_ME, LV_MS blocks

   python train_UNet.py
   

   • Train Mask-RCNN for LV_ME and LV_MS blocks
     • Download pretrained mask_rcnn_coco.h5 at [here] and place it in the current folder.
     • Train Mask-RCNN for LV_ME block

     python train_MaskRCNN.py --mode 'LV_ME'
     

     • Train Mask-RCNN for LV_MS block

     python train_MaskRCNN.py --mode 'LV_MS'
     

   • Train UNet++ for LV_ME and LV_MS blocks

   python train_UNetplusplus.py
   

3. Test networks:

   • Test UNet

   python test_UNet.py
   

   • Test Mask-RCNN
     • Test LV_ME block

     python test_MaskRCNN.py --mode 'LV_ME'
     

     • Test LV_MS block

     python test_MaskRCNN.py --mode 'LV_MS'
     

   • Test UNet++

   python test_UNetplusplus.py
   

4. Post-processing and linear decoder:

python post_processing.py

Acknowledgement

1. Please cite the official Mask-RCNN and UNet++ implementations if you use them:
   • Mask-RCNN: https://github.com/matterport/Mask_RCNN
   • UNet++: https://github.com/MrGiovanni/UNetPlusPlus
2. The authors would wish to acknowledge Compute Canada for providing the computation resource.