The XVFA (explainable vertebral fracture assessment) from the paper Explainable vertebral fracture analysis with uncertainty estimation using differentiable rule-based classification is an automated deep learning model for the detection and classification of vertebral fractures in lateral spine radiographs. The model is based on a differentiable rule-based classification approach that allows for the generation of explanations for the model's predictions. The model also provides uncertainty estimates for its predictions, which can be used to assess the reliability of the model's predictions.

The paper was accepted at the MICCAI 2024 conference, but a preprint version is available on arXiv.

The XVFA model is implemented in Python using the PyTorch and Lightning libraries. Clone this repository to your local machine and install the required dependencies using the following command:

pip install -r requirements.txt

The project depends on the DETR-DINO repository.

cd models/backbones
git clone https://github.com/IDEA-Research/DINO.git
cd DINO

pip install -r requirements.txt

cd models/dino/ops
python setup.py build install
# unit test (should see all checking is True)
python test.py
cd ../../..

The model can easily be trained on new data using the Lightning framework, by providing a Lightning datamodule adhering to the BaseDataModule class in xvfa/datamodules/base_datamodule.py. See the Lightning documentation for more information on how to create a custom datamodule.

All models may also be trained in a conventional PyTorch training loop, although examples of this are not provided in this repository.

If you use the XVFA model in your research, please cite the following paper:

@article{,
  title={Explainable vertebral fracture analysis with uncertainty estimation using differentiable rule-based classification},
  author={V. Wåhlstrand Skärström, L. Johansson, J. Alvén, M. Lorentzon and I. Häggström},
  journal={Lecture Notes in Computer Science, Medical Image Computing and Computer-Assisted Intervention – MICCAI 2024},
  year={2024},
  volume={},
  number={},
  pages={},
  doi={}
}