Building a deeplabv3 segmentation model - cerr/CERR GitHub Wiki

User-defined inputs

  • Directory containing DICOM images and segmentations (one sub-directory per patient).
  • JSON configuration file defining structure names, data split, hyperparameters etc.
    Sample : CERR/CERR_core/DLSegmentationTraining/sample_train_params.json

Exporting to .h5

  • DICOM data is converted to .h5 format required to train the deep learning model.

  • Example call:

prepareSegDataset(paramFilename);

Building a model

Requirements:

  • Pytorch
    Follow instructions for installation available at https://pytorch.org/

  • DeeplabV3
    Pull the PyTorch implementation of DeepLab-V3-Plus
    git pull https://github.com/jfzhang95/pytorch-deeplab-xception.git

  • Customize wrapper functions

    • mypath.py : Provide paths to .h5 datasets (training, validation and test directories)
    • dataloaders/init.py : Define dataloaders for the input dataset
    • dataloaders/utils.py : Map labels to colors for display of segmentation masks
    • datasets/yourscript.py : Define valid classes and image transformations