Implementation from the paper "Data Augmentation using Learned Transforms for One-shot Medical Image Segmentation".

Paper: arXiv link

This project has dependencies on the following repos. Please place all these repos in the same parent directory.

• https://github.com/xamyzhao/evolving_wilds (for general utils)
• https://github.com/voxelmorph/voxelmorph (for networks and losses)
• https://github.com/adalca/neuron (for SpatialTransformer layer)
• https://github.com/adalca/medipy-lib (for CPU implementation of Dice score)
• https://github.com/adalca/pynd-lib (for visualizations)

Spatial and appearance transform models can be trained by specifying the GPU ID, dataset name, and model name.

python main.py trans --gpu 0 --data mri-100-csts2 --model flow-bds
python main.py trans --gpu 0 --data mri-100-csts2 --model color-unet

Each experiment will create a results directory under ./experiments by default, so make sure that location exists.

A segmentation network can be trained with the following:

python main.py fss --gpu 0 --data mri-100-csts2

Again, results will be placed under .experiments. To evaluate trained segmenters, look at the code in evaluate_segmenters.py. You will have to modify the code to point at your trained models.

_{Repo name inspired by Magic: The Gathering.}