A implementation of the multi-level attention network for retinal vessel segmentation.
Before training, the images of training datasets are pre-processed with the following transformations:
- Gray-scale conversion
- Standardization
- Contrast-limited adaptive histogram equalization (CLAHE)
- Gamma adjustment
Detailed implementation is showed in ./lib/pre_processing.py
For the three databases, various transformations including flipping and rotation are applied on each image in the training set to generate 10 augmented images.
The training of the neural network is performed on sub-images (patches) of the pre-processed full images. Each patch, of dimension 256x256, is obtained by randomly selecting its center inside the full image. For the DRIVE database, a set of 4000 patches is obtained by randomly extracting 20 patches in each of the 200 augmented DRIVE training images. For the STARE database, a set of 5700 patches is obtained by randomly extracting 30 patches in each of the 190 augmented STARE training images (adopt the ’leave-one-out’ technique ). For the CHASE_DB1 database, a set of 8000 patches is obtained by randomly extracting 40 patches in each of the 200 augmented CHASE_DB1 training images.
Detailed implementation is showed in ./lib/extract_patches.py
The code is written in Python, it is possible to replicate the experiment by following the guidelines below.
The neural network is developed with the PyTorch library, we refer to the PyTorch for the installation.
The following dependencies are needed:
- numpy >= 1.13.3
- PIL >=4.3.0
- opencv >=2.4.10
- h5py >=2.8.0
- ConfigParser >=3.5.0b2
- scikit-learn >= 0.20.0
It is convenient to create HDF5 datasets of the ground truth, masks and images for both training and testing. In the root folder, just run:
python prepare_datasets_DRIVE.py
The HDF5 datasets for training and testing will be created in the folder ./DRIVE_datasets_training_testing/.
Now we can configure the experiment. All the settings can be specified in the file configuration.txt.
After all the parameters have been configured, you can train the neural network with:
python ./src/train.py
The parameters for the testing can be specified again in the configuration.txt file.
Run testing by:
python test.py
The following files will be saved in the folder:
- The ROC curve (png)
- The Precision-recall curve (png)
- Picture of all the testing pre-processed images (png)
- Picture of all the corresponding segmentation ground truth (png)
- Picture of all the corresponding segmentation predictions (png)
- One or more pictures including (top to bottom): original pre-processed image, ground truth, prediction
- Report on the performance
use the Wilcoxon’s rank-sum test [1] to conduct the significance analysis
python wilcoxon_test.py
[1] Mann, Henry B., and Donald R. Whitney. "On a test of whether one of two random variables is stochastically larger than the other." The annals of mathematical statistics (1947): 50-60.