Open-source domain adaptation to handle data shift for volumetric segmentation -use-case kidney segmentation”

Repository for benchmarking code for domain adaptation of 3D segmentation models

installl

• use a new conda env with python 3.10
• git clone the repo and cd into it.
• run pip install -e .

Dataset setup

• the notebook in ./domainadapt_segmentation/notebooks/step0_prep_datasets.ipynb contains descriptions for how to process the datasets
• It will specify how to create train,val and test sets

Training models

• Train models using
• python3 -m domainadapt_segmentation.train --config_path ./domainadapt_segmentation/example_configs/train_baseline.json
• You can specify your own configs in another directory

Testing Models

• How to test the model is also as simple as specifying a config file.
• An example is found in ./domainadapt_segmentation/example_configs/test_baseline.json. You must specify the following
  • model_weight: path to model checkpoint trained using our code.
  • output_dir: this is where the segmentations will be stored
  • metrics_path: this is a directory where we will store the disce score
  • device: wich GPU to run inference on
  • test set: the pickle file we will like to run inference on and evaluate performance

python3 -m domainadapt_segmentation.test --config_path ./path_to_test.json

Running Inference Only

• Fill out the parameters in example_configs/infer_example.json
• The parameters are as follows:
  • model_weight: path to weights. Two model weights are available at: gdrive
  • output_dir: directory to save output_segmentations
  • mapping_path: a csv that will give you two columns. path_of_input, path_output_seg
  • pkl_path: pickle path to test data organized like bellow
  • trainer: DiceTrainer. Jus specifies the inference logic.

python3 -m domainadapt_segmentation.infer --config_path ./path_to_infer.json

What are the pickle files

• We use monai dataloaders which expect the data to be given as a list of dictionarities
• We have pkl_contents= (train_set,val_set,test_set)
• Each of the train test splits is a list of dictionaries

train_set = [
    {'image':PathToVolume.nii.gz,
    'label':PathToMaskVolume.nii.gz,
    'phase': 1 if contrast else 0,
    'PatientID':Useful For stat calculation latet 
    }
    ]