Code for our MICCAI 2021 paper: Patch-free 3D Medical Image Segmentation Driven by Super-Resolution Technique and Self-Supervised Guidance
This is an end-to-end patch-free segmentation framework realizing HR segmentation with LR input. Our method uses Multi-task Learning as fundation and involves TFM and SGM to help HR restoration. SGM uses a representative HR patch to keep some HR information, and similar to multitask-learning, SGMs for US and SR also share the weights, in order for the segmentation task benefiting from the Self-Supervised Guidance Loss. TFM integrate the two tasks together and exploite the inter connections between the two tasks. In our experiments, our method outperforms the patch-based methods as well as has a four times higher inference speed.
Links to related further works:
Please follow the instructions on http://braintumorsegmentation.org/ to get your copy of the BRATS2020 dataset.
Our code should work with Python>=3.5 and PyTorch >=0.4.1.
Please make sure you also have the following libraries installed on your machine:
- PyTorch
- NumPy
- MedPy
- tqdm
Optional libraries (they are only needed when run with the -v flag):
- opencv-python
Firstly, clone our code by running
git clone git@github.com:Dootmaan/PFSeg.git
Normally you only have to specify the path to BRATS2020 dataset to run the code.
For example you can use the following command (of course you need to change directory to ./PFSeg first):
CUDA_VISIBLE_DEVICES=0 nohup python3 -u train_PFSeg.py -dataset_path "/home/somebody/BRATS2020/" > train_PFSeg.log 2>&1 &
You can also add the -v flag to have verbose output. Our code supports multi-gpu environment and all you have to do is specifying the indices of the available GPUs. Directly running train_PFSeg.py will use all the GPUs on your machine and use the default parameter settings. The minimun requirements for running our code is a single GPU with 11G video memory.
Click here to download the pretrained weights for our framework.
Please note that the code uses 6/2/2 train/val/test split by default, while our results in the paper are reported with a 8/2 train/test split so if you would like to verify the results please make sure you change the default dataset split. We also did a 5-fold cross validation for our method and the results are quite stable, as you can see in the rebuttal part of our paper.
More codes and information may be updated later.
Below are some experimental results that may be helpful (BRATS2020).
Results with 8/2 train/test split
Method | DSC | HD95(mm) |
---|---|---|
VNet | 0.7991(0.1345) | 13.86(18.6006) |
UNet3D | 0.8121(0.1277) | 14.63(21.1204) |
ResUNet3D | 0.8218(0.1182) | 13.21(19.4195) |
ResUNet3D⬆ | 0.8089(0.1525) | 8.56(8.0853) |
Holistic Decomposition+ResUNet3D | 0.8245(0.1474) | 9.21(12.0561) |
Ours | 0.8382(0.1433) | 7.83(8.6250) |
Results with 6/2/2 train/val/test split
Method | DSC | HD95(mm) |
---|---|---|
VNet | 0.7776(0.1693) | - |
UNet3D | 0.7954(0.1368) | - |
ResUNet3D | 0.8097(0.1291) | - |
ResUNet3D⬆ | 0.8016(0.1595) | - |
Holistic Decomposition+ResUNet3D | 0.8161(0.1645) | - |
Ours | 0.8329(0.1450) | 8.5131 |
6/2/2 Ablation Study:
Method | UNet | ResUNet |
---|---|---|
US | 0.7945 | 0.8082 |
US+SR | 0.8010 | 0.8141 |
US+SR+TEL | 0.8106 | 0.8193 |
US+SR+TEL+SSL | 0.8163 | 0.8281 |
US+SR+TEL+SSL+SGM | 0.8208 | 0.8329 |