Transmutat-PDXA is a generative deep learning model which produces qualitatively and quantitatively accurate dual energy X-ray absorptiometry (DXA) scans from three-dimensional (3D) body surface scans. This groundbreaking medical imaging model demonstrates quantitative precision on commercial clinical software, yielding accurate measures of fat mass, lean mass, and bone.
Transmutat-PDXA runs on Linux and requires a GPU.
Install dependencies:
- Python >= 3.6
- tensorflow >= 2.4
- plyfile >= 0.7.4
- opencv-python >= 4.5.2
- numpy >= 1.19.5
All 3D scan meshes must be fitted and standardized to a T-pose with 110K points using the Meshcapade API.
Update the config file:
- GPU device ID. Variable = [setup][gpu]
- Path to model weights. Variable = [model][transmutat_pdxa_weights]
- Path to ply file or 3D mesh. Variable = [predict][mesh_path]
- Path to output dir. Variable = [predict][out_path]
Weights for a version of the Transmutat-PDXA model and a fitted mesh for testing are available upon request via the SRL Website.
python predict.py -c model.config
Expected outputs: Raw DXA as a 6-channel npy file, raw DXA single-channel PNG, and a processed DXA single-channel PNG.
Code for transforming NPY files into proprietary DXA files for analysis cannot be made publicly available. Those at SRL can perform the transformation and should be contacted accordingly. The resulting DXA file can then be analyzed using DXA body composition analysis software as shown below.
The base model was trained using self-supervised learning on over 20K DXA scans from the Bone Mineral Density in Childhood Study and The Health, Aging and Body Composition Study. Code for the self-supervised model is available here: DXA-VAE.
Training of the final model was conducted on patients from the Shape Up! Studies, who received paired DXAs and 3D scans on the same day. The training schema is shown below.
For questions or more information, please reach out via email: contact@transmutat.ai