How uncbiag/unigradicon works
Project type and purpose
• This is a Python-based full-stack research tool. It ships as a library (src/unigradicon), a set of CLI executables (unigradicon-register, unigradicon-warp), training scripts (training/*.py), and a Slicer extension. Its purpose is to perform medical image registration via pre-trained foundation models (uniGradICON, multiGradICON) and to fine-tune/train new models.
File structure
• Top-level:
• src/unigradicon/ – the core package containing submodules for commands, model definition, IO, loss, utilities.
• training/ – dataset loaders (dataset.py, dataset_multi.py) and training entrypoints (train.py, train_multi.py).
• demos/ – usage examples and figures.
• tests/ – unit tests covering CLI argument parsing, ITK interface, requirement syncing.
• Project metadata (pyproject.toml, requirements.txt, setup.cfg) and CI workflows under .github/workflows.
README for additional insights
• It describes three main usage modes: CLI, Colab notebook, Slicer extension.
• Internally, images and transforms are handled via ITK; models are PyTorch networks; weights are auto-downloaded from GitHub.
• The registration pipeline consists of model inference plus optional instance-optimization (IO) loops with configurable similarity measures and loss function masking.