3D Efficient Subpixel-shifted Convolutional Network (3D-ESPCN)

This chainer implementation is based on "Tanno R. et al. (2017) Bayesian Image Quality Transfer with CNNs: Exploring Uncertainty in dMRI Super-Resolution. In: Descoteaux M., Maier-Hein L., Franz A., Jannin P., Collins D., Duchesne S. (eds) Medical Image Computing and Computer Assisted Intervention − MICCAI 2017. Lecture Notes in Computer Science, vol 10433. Springer, Cham".

Note that this is not official implementation.

The difference between original paper and this as follow:

• Dataset

  • I used Balloon Analog Risk-taking Task dataset

• I only implement baseline model (3D-ESPCN).

  • Network architecture

    

• Definition of pixel shuffler

  I think this is correct definition.

  

  F: input feature map

  c: number of output image channel

  S: Pixel shuffler

  i, j, k, c: coordinate in output image

  r: upsampling rate

Requirements

• chainer

• cupy

• SimpleITK

• pyyaml

How to use

1. Download dataset here.

   Please put all dataset to data/raw after you unzipped it.

2. Make mhd data and LR image

   # Make mhd data in data/interim
   python util\miscs\clean_data.py
   
   # Make LR and HR images in data/processed
   python util\miscs\make_lr_img.py
   

   • LR image sample (x1/4)

   

   • HR image sample

   

3. Train model

   python training.py -g 0
   

   • Training result

     

4. Infer HR images and evaluate infered HR images in terms of PSNR and SSIM.

   python inference.py -g 0 -m results\training\gen_iter_100000.npz
   

   • Inference results

     Case	PSNR[dB]	SSIM
     0014.mhd	29.4	0.789
     0001.mhd	30.0	0.767
     0015.mhd	34.7	0.852

   • Example of ground truth (0015.mhd)

     

   • Example of input LR image (0015.mhd)

     

   • Example of infered HR image (0015.mhd)

     

My opinion

• I think I cant reconstruct detail of image, e.g. texture 😭

• If you use this implementation, you should optimize this model for your task.

• Note that this is just my hobby. So, please dont care this results. 😸