panakino / fbpconv_tf

tensorflow implementation for FBPConvNet

FBPConvNet - tensorflow

http://ieeexplore.ieee.org/document/7949028/

This is tensorflow implementation for Deep Convolutional Neural Network for Inverse Problems in Imaging, TIP (2017).

• applications (forward model with shift-invariant normal operator):

• 2D sparse-view CT reconstruction
• reconstruction of accelerated MRI
• Deconvolution of shift-invariant

Whole codes are forked and modified from https://github.com/jakeret/tf_unet.

Training configuration

• Tensorflow 1.1.0
• 1 or 2 GPUs (TITAN X pascal arch.)
• MacOS X 10.12.6
• Python 2.7.12

Data - XYCN format (ellipsoids, downsampling factor : x 20)

• train : https://drive.google.com/open?id=1FTOgM2vOQaGSokEDtOaPNdBTto6h5yFi
• test : https://drive.google.com/open?id=1w_kPao6L2UwhTKIgcr_3o62A6vYYtX_r
• If you want to make fbp images, you can find file_generator in tf_unet/layers.py (load_whole_data function.)

illustration

Commands

Before starting,

pip install pillow matplotlib scipy scikit-image h5py

To start training a model for FBPConvNet:

python main.py --lr=1e-4 --output_path='logs/' --train_path='train_data/*.mat' --test_path='test_data/*.mat' --features_root=32 --layers=5 

To deploy trained model:

python main.py --lr=1e-4 --output_path='logs/' --train_path='train_data/*.mat' --test_path='test_data/*.mat' --features_root=32 --layers=5 --is_training=False

You may find more details in main.py.

Contact

[email protected]

fbpconv_tf's Issues

how to use load_whole_data function to get FBP images?

As I want to make fbp images, I want to have a sight of what 'load_whole_data function' can do.
As the dataset provided is downsampling ellipsoids images after FBP process, what data can be fed to 'load_whole_data function' ?

Thanks a lot!

image_gen some code is skipped..

explain whole process

sir, could you please explain how the whole process goes, how it starts, and process
I'm confusing with the given source please clarify me
thank you