Mining the manifolds of deep generative models for multiple data-consistent solutions of tomographic imaging problems - PyTorch implementation

Paper: https://arxiv.org/abs/2202.05311

Sayantan Bhadra¹, Umberto Villa¹ and Mark A. Anastasio^2
1Washington University in St. Louis, USA
²University of Illinois at Urbana-Champaign, USA

---

System requirements

• Linux
• Anaconda >= 2018.2
• Python >= 3.6
• NVIDIA GPU(s) (compute capability GeForce GTX 1080 or higher and minimum 8 GB RAM)
• NVIDIA driver >= 440.59, CUDA toolkit >= 10.0

Environments

• PULSE++ and PULSE: Create a virtual environment with the following command:

conda create -f pulse_pp.yml

• Langevin sampling: Set up virtual environment by running the commands at https://github.com/utcsilab/csgm-mri-langevin#setup-environment, with root directory as MRI/langevin/

Downloads

Download pre-trained model weights of StyleGANs and NCSNv2, and the system matrix for the limited-angle CT imaging system from UIUC Databank: https://doi.org/10.13012/B2IDB-8852490_V1

Undersampled k-space experiments

The scripts for PULSE++ and PULSE should be run with root directory as MRI/. For running Langevin sampling, the root directory should be MRI/langevin.

Data subdirectories in MRI:

• knee_1 and knee_2: Folders containing objects Knee 1 and Knee 2
• kspace_1 and kspace_2: Folders containing undersampled k-space data from Knee 1 and Knee 2
• masks: Folder containing 6-fold and 8-fold random Cartesian sampling masks

PULSE++

1. Save the downloaded MRI-StyleGAN model weights file MRI_synthesis.pt under MRI/ (used for both PULSE++ and PULSE)
2. Run ./run_meas_alpha.sh

PULSE

1. Run ./run_meas.sh

Langevin sampling

1. Go to the directory MRI/langevin
2. Save the downloaded MRI-NCSNv2 model weights file MRI_ncsnv2.pth under MRI/ (used for both PULSE++ and PULSE)
3. Run ./run_multiple_no_norm.sh

Limited-angle CT experiments

Run PULSE++ with root directory as CT/

Data subdirectories in CT:

• lung_1 and lung_2: Folders containing objects Lung 1 and Lung 2
• proj_1 and proj_2: Folders containing limited-angle projection data from Lung 1 and Lung 2

PULSE++

1. Place the downloaded system matrix CT_H_matrix.mat in the directory CT/
2. Run run_meas_alpha_airt_kl.sh