Neural Fields for Tomographic Reconstruction

Samuele Papa, David M. Knigge, Riccardo Valperga, Miltiadis Kofina, Nikita Moriakov, Jan-Jakob Sonke, Efstratios Gavves

In this project we improve neural field methods to quickly perform tomographic reconstruction. Special care has been put in making realistic assumptions and applying the method to a large open-source dataset.

Installation

The easiest way to run the code is to create a conda environment and install the dependencies from the environment.yml file.

conda env create -f environment.yml

This will create a new environment called tomorecon.

After the creation of this environment a custom package provided with this repository must be installed to process the geometry files used to interpret the projections correctly. Use the following to perform the installation (the -e option is used to install the package in editable mode):

conda activate tomorecon
cd tomo_utils_package
pip install -e .

Experiments

Before running any experiment using the conditioning method, use the following to setup the data in the correct format:

python data_factory/make_recon_data.py --volumes_dir=path_to_original_volumes --recons_dir=path_to_recon_volumes

To reproduce the training of the model used in this paper, use the following command:

python experiments/train_validate.py run_name=train project_name=reproduce dataset.path=path_to_recon_volumes dataset.original_volumes_path=path_to_original_volumes val_dataset.path=path_to_original_volumes dataset.name=reconstructions val_dataset.name=projection nef.type=Hash nef.num_hidden=128 nef.num_layers=2 nef.final_act=relu training.lr=5e-05 nef.conditioning.spatial_cfg.type=Hash nef.conditioning.spatial_cfg.num_layers=2 nef.conditioning.spatial_cfg.num_hidden=128 training.max_time=3600000000 dataset.num_vols=200 hydra.run.dir=path_to_output