Preprocess and train scripts for complete mitral valve segmentation from ultrasound data
Place previously extracted train data into the raw\train directory. Each .nrrd image needs a corresponding
mask with the same filename and a _mask suffix. For example train_image01.nrrd would need a mask file named
train_image01_mask.nrrd. Otherwise the data process script cannot load the mask data.
To use the broken data folder place the broken mask files into subdirectories under the raw/broken directory.
The subdirectories need the class of data quality as folder name. For exmaple raw/broken/6 has the broken
files with six seperate parts of the mitral valve. Also make sure cleanup_broken_data() is called in the
main function before loading the data.
The images for testing should be in the raw/test directory. After train a prediction for the mask on each image
is made.
If the data is placed into the right folders, just run the data.py script with Python and watch
how the numpy files are generated (in case enough space is available).
After the generation of the numpy files just run the train.py script. Training will start, default 20 epochs.
This parameter can be easily adjusted in the code. Just check the parameters of the model.fit function.
The trained weights are used after the training to predict masks on the loaded test files. The output includes .nrrd and .tiff files for more convenient data access. You can find them in the preds directory.
raw/
├── train
├── test
└── broken