S-cuda: Self-Cleansing Unsupervised Domain Adaptation for Medical Image Segmentation
This repository provides code for the paper, S-CUDA: Self-Cleansing Unsupervised Domain Adaptation for Medical Image Segmentation. Please refer to our paper for better understanding the method. Notably, the N1 or N2 in the framework can be replaced to other segmentation networks based on your tasks.
Pipeline
Getting started
Environments
- python 3.5
- tensorflow 1.4.0
- keras 2.2.0
- pytorch 0.4.0
- CUDA 9.2
Packages
- tqdm
- skimage
- opencv
- scipy
- matplotlib
Datasets
Download from Refuge, prepare dataset in data directory as follows.
S-cuda
│ Network-1
| Network-2
│ scripts
└───dataset
│ │ source
│ │ │ images
│ │ │ labels
│ │
│ └───target
│ │ │ images
│ │ │ labels
│ │ │ pseudo_label
│ │
│ └───test
│ │ │ images
│ │ │ labels
│ │
│ └───source.txt
│ │ target.txt
│ │ test.txt
│
└───README.md
Initial weights and pre-trained model
Initial weights and pre-trained model download link:
- Initial weights, code:4y69
- Pre-trained model, code:9koj
unzip Initial_weights.zip
unzip Pre-trained model.zip
Running
0.Clone this repo:
git clone https://github.com/zzdxjtu/S-cuda.git
cd S-cuda
1.Train:
All training script is stored in scripts directory.
sh scripts/run1.sh
##The two peer networks select the noise data and clean data simultaneously, and then the selected clean data are fed into the peer network for finetune, and you can change the remember rate and noise rate according to the noise ratio of the source training set.
sh scripts/run2.sh
##The two networks select the noise data and clean data simultaneously, and then input the clean data into each other's network for finetune.
sh scripts/run3.sh
##Correct the common part of the noise data selected by the two networks, and then the two networks select the noise data and clean data simultaneously, at last input the clean data into each other's network for finetune.
sh scripts/run4.sh
##Correct the common part of the noise data selected by the two networks, and then the two networks select the noise data and clean data simultaneously, at last input the clean data into each other's network for finetune.
Before training, please check whether all the model weight and dataset path is correct.
2.Test:
cd S-cuda
python Network-1/evaluation.py
3.Performance
Pretrained
| Level_0.5-0.7/noise_labels_0.5 | select_0.1 | select_0.2 | select_0.3 | select_0.4 |
|---|---|---|---|---|
| Disc_dice | 0.95 | 0.95 | 0.948 | 0.95 |
| Cup_dice | 0.882 | 0.873 | 0.871 | 0.871 |
Scratch
| Level_0.5-0.7/noise_labels_0.5 | select_0.1 | select_0.2 | select_0.3 | select_0.4 |
|---|---|---|---|---|
| Disc_dice | 0.947 | 0.943 | 0.943 | 0.941 |
| Cup_dice | 0.889 | 0.886 | 0.886 | 0.872 |
4.Selection ability
We also demonstrate the selection ability of these two networks training from scratch via calculating the dice coefficient.
Clean data
| N1/Level_0.5-0.7/noise_labels_0.1 | select_0.1 | select_0.3 | select_0.5 | select_0.7 |
|---|---|---|---|---|
| Disc_dice | 1.0 | 0.995 | 0.997 | 0.994 |
| Cup_dice | 1.0 | 1.0 | 1.0 | 0.995 |
| N2/Level_0.5-0.7/noise_labels_0.1 | select_0.1 | select_0.3 | select_0.5 | select_0.7 |
|---|---|---|---|---|
| Disc_dice | 1.0 | 1.0 | 0.995 | 0.983 |
| Cup_dice | 1.0 | 1.0 | 0.982 | 0.967 |
Noisy data
| N1/Level_0.5-0.7/noise_labels_0.1 | select_0.1_1 | select_0.1_2 | select_0.1_3 | select_0.1_4 |
|---|---|---|---|---|
| Disc_dice | 0.861 | 0.896 | 0.786 | 0.815 |
| Cup_dice | 0.784 | 0.777 | 0.755 | 0.813 |
| N2/Level_0.5-0.7/noise_labels_0.1 | select_0.1_1 | select_0.1_2 | select_0.1_3 | select_0.1_4 |
|---|---|---|---|---|
| Disc_dice | 0.857 | 0.864 | 0.833 | 0.897 |
| Cup_dice | 0.766 | 0.801 | 0.827 | 0.925 |
| Intersection of N1 and N2/Level_0.5-0.7/noise_labels_0.1 | select_0.1_1 | select_0.1_2 | select_0.1_3 | select_0.1_4 |
|---|---|---|---|---|
| Disc_dice | 0.761 | 0.831 | 0.7 | 0.743 |
| Cup_dice | 0.615 | 0.596 | 0.715 | 0.866 |
In addition, the noise select ratio and clean select ratio can be modified according to your own tasks and the training dataset. For example, with the noise ratio of the dataset, which can be estimated by sampling analysis, you can set the select ratio accordingly. You can also set both of the initial noise select ratio and clean select ratio to 0.1, respectively, and then increase the select ratio of clean data gradually to make sure the two peer netowrks work well, after that maintain the select ratio of clean data and gradually increase the select ratio of noisy data.
Supplementary notes
boudary.ipynb
##Calculate the weight map of the optic cup, optic disc, and background
calculate_dice.py
##Calculate dice coefficient
get_contour.ipynb
##Obtain the edge contour of the target object
hausdorff_dis.py
##Calculate hausdorff distance
noise_label.ipynb
##Generate labels with different levels of noise and different proportions, including corrosion, expansion, deformation operations
crop.py
##crop 512×512 ROI from source image
Acknowledge
Some of our codes (i.e., two peer networks) are referring to liyunsheng13/BDL and EmmaW8/pOSAL. Thanks for their helpful works.