PVCNet

A Deep Learning-based Partial Volume Effect correction method, for Single Photon Emission Computed Tomography (SPECT).

If you use this in your research, please cite:

  @article{kaprelian2025PVCNet,
  	title={Partial Volume Correction of SPECT Images With Deep Learning Trained on Synthetic Data},
  	author={Kaprélian, Théo and Etxebeste, Ane and Sarrut, David},
  	journal={IEEE Transactions on Radiation and Plasma Medical Sciences}, 
  	year={2025},
	doi={10.1109/TRPMS.2025.3634484}
	}

Training Dataset

To generate a synthetic dataset of Lu177 sinograms, refer to this repository

The generated dataset should be in .h5 format. Each sample should be a five letter keys (like "ABCDE"). For each sample the training dataset should contain the following keys:

Sinogram-based training (PVCNet-sino)

• "PVfree_att" key containing the ground truth sinogram without PVE (no collimator-detector modeled) obtained with attenuated forward-projection
• "PVE_att"
• "PVE_noisy"
• "rec_fp_att"
• "attmap_fp"

Image-based training (PVCNet-img / PVCNet-hybrid)

• "rec"
• "attmap_4mm"
• "PVCNet_rec"

Training

To train a PVCNet, use this command:

python train.py --json configs/config_file.json --output EXAMPLE_REF --output_folder /output/folder/

Two main types of config files:

Sinogram-based training (PVCNet-sino)

Use a file similar to configs/config_PVCNet_sino.json, i.e.:

• with a dataset/test_dataset containing the sinogram keys (PVfree_att, PVE_att, PVE_noisy, rec_fp_att, attmap_fp)

Two choices:

• Either with 2 NN (one for denoising, one for PVC). Set

  "network": "unet_denoiser_pvc",
  

• Either with 1 NN. Set

  "network": "unet",
  

Image-based training (PVCNet-img / PVCNet-hybrid)

Use a config file similar to configs/config_PVCNet_img.json.

Once again two choices:

• Use only "rec" (the reconstructed image with RM) and the attenuation map as input, then set

  "with_PVCNet_rec": false
  

This is the PVCNet-img.

• Also use "PVCNet_rec" as input, then set

  "with_PVCNet_rec": true
  

Then, it is the PVCNet-hybrid because it uses the image reconstructed from the sinogram corrected by a previously trained PVCNet-sino.

Apply

For PVCNet-sino

 python apply.py --input projections.mha --input_rec_fp projections_rec_fp.mha \\
  --attmap_fp attmap_fp.mha --pth path/to/network.pth\\
  --output output_projections.mha

For PVCNet-img

python apply.py --input rec.mhd --attmap_fp attmap_4mm.mha \\
 --pth path/to/network.pth\\
 --output output_image.mha 

For PVCNet-hybrid

python apply.py --input rec.mhd --input_rec_fp rec_noRM_PVCNet_sino_751113.mhd\\
 --attmap_fp attmap_4mm.mha --pth path/to/network.pth\\
 --output output_image.mha 

Trained models

Examples of trained models can be found in the "pth" folder