Paper Accepted at IEEE EMBC 2022

You can find all DRR in the following link. Here is a description of the folders:

An instance comprehends 72 DRRs (each at 5 degrees) from a 360 degree rotation of a real CT scan.

chest_xrays all images of the 20 chest instances (.png, res. 128x128).

knee_xrays all images of the 5 knee instances (.png, res. 128x128)

You can download our models' weights through the following links:

Knee model

Chest model

Refer to graf-main folder and execute, replacing CONFIG.yaml with knee.yaml or chest.yaml

python train.py configs/CONFIG.yaml

After training a model, you can test its capacity to reconstruct 3D-aware CT projections given a single X-ray.

To execute the reconstruction, please refer to graf-main folder and execute:

python render_xray_G.py configs/experiment.yaml /
    --xray_img_path path_to_xray_folder /
    --save_dir ./renderings /
    --model path_to_trained_model/model_best.pt /
    --save_every 25 /
    --psnr_stop 25 

Update: This step requires more memory as rays are now tracked during the optimization process, so the higher the resolution of the image, more memory it requires. For this, we add an additional argument (img_size) with a default value of 64 if GPU size <=10GB.

Optimizing both G and z (render_xray_G_Z.py): By setting the model to generator.train() when calling reconstruction(), we can optimize z as well and NLL will change over time. Finetuning like this, optimizes faster when considering a single X-ray projection, however, G would need to figure out 3D consistency for this new instance almost from scratch.

Optimizing G only (render_xray_G.py): By setting generator.eval(), only G and part of its parameters will be optimized. NLL doesn't play a role anymore in the reconstruction process. 3D consistency from G is preserved.

We consider optimizing G only the way to go (at least for now). This open new future experiments. Both rendering scripts if you'd like to try them.

First, download the camera poses from this link. We referred to Plastimatch's projection geometry and Projection matrix documentation to generated the camera matrices. To use pixelNeRF model download the following configuration files here and place them in the following folders:

pixel-nerf/conf/exp/ct_single.conf
pixel-nerf/conf/exp/drr.conf

Also refer to the following script to render CT images:

pixel-nerf/eval/render_ct.py

To generate xrays images (.png) at different angles from CT scans use the script generate_drr.py under the folder data/. To run it you need to install the Plastimatch's build. Version 1.9.3 was used.

An updated version of the script has been added (generate_drr_multiple_dirs.py). Use this script to automatically generate the set of DRRs when you have a global folder with multiple folders containing CT scans. The files do not necessarily need to be in the immediate subfolder. You only need to assign the path location of the global folder and the global folder to save the set of DRRs.

Replace the following variables within the file:

• input_path: path to the .dcm files or .mha file of the CT.
• save_root_path: path where you want the xrays images to be saved.
• plasti_path: path of the build.
• multiple_view_mode <True | False>: generate single xrays from lateral or frontal views or multiple images from a circular rotation around the z axis. If False you need to specify the view with the argument frontal_dir <True | False> (false for lateral view). If True you need to specify num_xrays to generate equally spaced number of views and angles to input the difference between neighboring angles (in degrees).
• preprocessing <True | False>: set this to True if files are .dcm for Hounsfield Units conversion. Set to False if given file is raw (.mha), for which you need to provide its path under the variable raw_input_file.
• detector_size: pair of values in mm
• bg_color: choose either black or white background.
• resolution: size of the output xrays images.

This codebase is heavily based on the GRAF code base. We also use the code from pixel-nerf for baseline experiments.

We thank all authors for the wonderful code!

If you use our model for your research, please cite the following work.

@misc{coronafigueroa2022mednerf,
      title={MedNeRF: Medical Neural Radiance Fields for Reconstructing 3D-aware CT-Projections from a Single X-ray}, 
      author={Abril Corona-Figueroa and Jonathan Frawley and Sam Bond-Taylor and Sarath Bethapudi and Hubert P. H. Shum and Chris G. Willcocks},
      year={2022},
      eprint={2202.01020},
      archivePrefix={arXiv},
      primaryClass={eess.IV}
}