Skip to content

Latest commit

 

History

History
30 lines (21 loc) · 1.53 KB

readme.md

File metadata and controls

30 lines (21 loc) · 1.53 KB
Raw

Image reconstruction with priors

Congrats 😃, you get to the final step. Before performing the reconstruction task, make sure you have installed bart toolbox properly with TensorFlow graph support. We prepare a tutorial to present how to create a regularization term with TensorFlow and use it for image reconstruction in BART.

If you'd like to have a quick tryout with our priors, you could get priors from this huggingface page or zenodo

We have two main steps to improve reconstruction with your priors. Firstly, export your trained model as computation graph for BART; then, use pics or nlinv command in BART with the exported graph as regularization. We provide an example for 2D reconstruction.

...

## export graph
log=../../MRI-Image-Priors/PixelCNN/cplx_large  # the folder that has models
meta=pixelcnn  # the name of a model
path=../../MRI-Image-Priors/exported/graph
name=pixelcnn_cplx_large
python 2d_pixelcnn.py $log $meta $path $name PixelCNN none 2DCPLX
GRAPH=$path/$name

...
...

## perform reconstruction
bart pics -g -i100 -d4 -R TF:{$GRAPH}:0.8 und_kspace coilsen prior_abide_pics

Homework

If you are interested in deep experience with our prior, you can generate a raw 3D k-space with sequence t1_tfl3d_ns_sag_TI900_FA9 on a SIEMENS scanner and reconstruct image from it with tfl3d.sh.