DXA-VAE

This repo houses code used for NeurIPS 2022 submission entitled "Quantitative Imaging Principles Improves Medical Image Learning"

• DXA = Dual Energy X-ray Absorptiometry
• VAE = Variational AutoEncoder

---

---

Prerequisites

• Linux Kernel: 4.4+
• CUDA Version: 10.1
• Python 3.6+ and the following modules:
  • Keras_Preprocessing==1.1.2
  • matplotlib==3.3.4
  • numpy==1.19.2
  • pandas==1.2.3
  • tensorflow==1.13.1

Usage

• In train.config file:
  • define path to directory containing data split csv files
    • directory should contain a files labeled "train.csv", "val.csv", and "test.csv"
      • csv files should contain one column and each row contains a path to DXA npy files

      paths
      path_to_image1
      path_to_image2
      ...

  • define model parameter
  • specify path for saved model
• run train.py

Example

python train.py -c train.config

---

Datasets

Data used for training and experiments are available upon request. Inquire at https://shepherdresearchlab.org/services/

---

Models and Weights

Trained VAE, encoder, and Pseudo-DXA models are available upon request. Inquire at https://shepherdresearchlab.org/services/

Citation

@article{leong2022quantitative,
  title={Quantitative Imaging Principles Improves Medical Image Learning},
  author={Leong, Lambert T and Wong, Michael C and Glaser, Yannik and Wolfgruber, Thomas and Heymsfield, Steven B and Sadowski, Peter and Shepherd, John A},
  journal={arXiv preprint arXiv:2206.06663},
  year={2022}
}