Enhancing Dynamic CT Image Reconstruction with Neural Fields and Motion Models

Take a look at the preprint here

The ground truth is two squares moving around some background.

Let us see some reconstructions!

Reconstruction for $\gamma=10^{-2}$ and full batch training $N_{mb}=N_T$

Reconstruction for $\gamma=10^{-2}$ and mini batch training $N_{mb}=1$

Reconstruction for $\gamma=0$ and full batch training $N_{mb}=N_T$

Reconstruction for the voxelated approach for $\alpha=\beta=\gamma=10^{-3}$.