msmt-incseg

Official implementation of the ECCV-MCV workshop submission "Multi-Scale Multi-Task distillation for incremental 3D medical image segmentation"

Reference

Part of the implementations were inspired by

• pytorch-3dunet
• UCD

Requirements

To install the backbone package pytorch3dunet:

pip install -e .

Additional requirements might include:

torch=1.7.1
nibabel
SimpleITK
scipy
numpy
h5py
matplotlib
seaborn
imgviz
skimage
labelme
opencv-python
Pillow

Data Download and Preprocess

NCI-ISBI2013 dataset

The NCI challenge dataset can be obtained from the official website and the download page

Place the download folders into a directory structured as /data/<USERNAME>/data/dynamic_segmentation/nci-isbi2013 (place your in it), then run the preprocessing code pytorch3dunet/datasets/preprocess_nci.py after properly replacing the desensitized token .

BraTS2015 dataset

The BraTS2015 dataset can be obtained from its challenge website

Place the download folders into a directory structured as /data/<USERNAME>/data/dynamic_segmentation/brats2015/train (place your in it), then run the preprocessing code pytorch3dunet/datasets/preprocess_brats_v2.py after properly replacing the desensitized token .

Training and Inference

Training scripts corresponding to Tables 1 and 2 in the paper are provided under scripts/

• For NCI-ISBI2013 dataset, look into scripts/eccv_nci/*.sh
• For BraTS2015 dataset, look into scripts/eccv_brats/*.sh

The file name is corresponding to items in the table, for example, to reproduce training for Mem2+MSMT for NCI-ISBI2013, please run scripts/eccv_nci/Mem2_MSMT.sh

For each of the script, please ensure to modify the following before executing:

• Replace all occurences of the <USERNAME> token with your specified value properly. You might need to create similar directory structures as specified in these scripts on your own system.
• Specify device=? in the script. If you have two gpu devices for example, device could be 0 or 1.

Generate tables in paper

After the training finished, you can run gen_paper.py to generate the running average dice numbers in the tables. Please properly specify meta information in paper_data.py.

Note that you also need to replace all occurences of the <USERNAME> token with your specified value properly in these codes.

Contact the author

Should there be any questions, please contact the author directly at the [email protected]