This repository contains the codebase aimed at 3D multi-organ segmentation in CT, using directed acyclic graphs and edge skip-connections.

The conference paper is "Graph-based Regional Feature Enhancing for Abdominal Multi-Organ Segmentation in CT."

Install the following essential packages using pip install command:

• torch==1.7.1 (Or install other versions of PyTorch corresponding to your CUDA versions using commands on the PyTorch website.)
• numpy==1.19.2
• MedPy==0.4.0

Store images, labels, and edge maps in the following folders:

/path/to/data/image/image0002.nii.gz

/path/to/data/label/label0002.nii.gz

/path/to/data/edge/edge0002.nii.gz

Make a look-up table using .jason format recording image, label and edge map directories of each patient, for example, our ./data/cross_validation.jason.

./cacheio/Dataset.py contains the class Dataset, which is a subclass of torch.utils.data.Dataset used for obtaining image, label, and edge map directories of a patient given the look-up table and then loading and augmenting data. Rewrite the class functions to make sure that data can be correctly loaded.

The process of cross-validation training is defined in ./train_full_scheme.py.

The option --fold=0 specifies the cross-validation fold, and --cv_json='/path/to/cross_validation.jason' the path of a look-up table indexing your data at hand.

Modify hyper-parameters if necessary which are by default --batch_size=2, --optim='adam', --lr=1e-3, --weight_decay=3e-4, --num_epoch=400, and --beta=1 and --beta2=1 specifying the weights of the segmentation and edge detection objectives respectively.

The evaluation process defined in ./inference.py computes the following volumetric metrics: the Dice similarity coefficient, the 95th percentile Hausdorff distance, and the average symmetric surface distance.

Specify the validation fold --fold and the path of a look-up table recoding directories of validation data --cv_json.