nn-extractor

Getting Started

If you have a trained nnUNetv2 model and the testing images, then you can extract the underlying information with:

# git clone and cd to the code directory.
git clone git@github.com:chhsiao1981/nn-extractor.git
cd nn-extractor

# install this package.
pip install -e .

# run scripts
./scripts/run-nnunetextractor.sh [nnUNet_rootdir] [nnUNet_train_dataset] [nnUNet_config (2d, 3d_fullres)] [nnUNet_fold (1,2,3,4,5,all)] [input_dir] [nnUNet_output_dir] [nn-extractor config]

# presented as web-api.
./scripts/dev_server.sh

example to run ./scripts/run-nnunetextractor.sh:

./scripts/run-nnunetextractor.sh /mnt/nnUNetv2 1 3d_fullres all /mnt/nnUNetv2/raw/Dataset001_BONBIDHIE/imagesTs /mnt/nnUNetv2/predicts/Dataset001_BONBIDHIE config.toml

Goal

Given a trained deep neural network model and inputs. nn-extractor extracts all the relevant information starting from inputs to output results. Currently we focus on 3D medical imaging segmentation tasks.

Structure of the Extraction.

Given a deep neural network model and an input, the extracted information can be divided as the following categories:

• Input: the inputs.
• Preprocess: preprocessing information.
• Forward: snaphots of forward prediction process.
• Backward: snapshots of backward propagation process.
• Postprocess: postprocessing information.
• Output: the outputs.
• (Sub-)extractor: The recursive sub-extractors for sub-tasks.
• Taskflow: The sequence of the task-flow.