hanyoseob/HDD-DL-for-SVCT

Hierarchical Decomposed Dual-domain Deep Learning for Sparse-View CT Reconstruction

Requirements

PLEASE INSTALL !!! DOCKER !!! TO CREATE CONTAINER.

https://docs.docker.com/engine/install/ubuntu/

Getting Started

1. Clone this repository

   git clone git@github.com:hanyoseob/HDD-DL-for-SVCT.git
   

2. Create, Start, and Run Docker Container

   docker run -it -v [HOST_DIR]:/workspace/[CONTAINER_DIR] --gpus all --name [DOCKER_NAME] pytorch/pytorch:1.11.0-cuda11.3-cudnn8-devel /bin/bash
   

   ex)

   docker run -it -v ./HDD-DL-for-SVCT.:/workspace/HDD-DL-for-SVCT. --gpus all --name mpi pytorch/pytorch:1.11.0-cuda11.3-cudnn8-devel /bin/bash
   

3. Install basic libraries

   apt-get update
   

   apt-get install git vim wget unzip -y
   

4. Move to the repository

   cd HDD-DL-for-SVCT.
   

5. Install the requirements

   pip install -r requirements.txt
   

6. Download pre-trained models

   wget https://www.dropbox.com/scl/fi/ryvda57j7dgu8yleqi51i/checkpoints.zip?rlkey=s6mczko9ajsm5va2x40e4u51t
   

   mv checkpoints.zip?rlkey=s6mczko9ajsm5va2x40e4u51t checkpoints.zip 
   

   unzip checkpoints.zip
   

   rm checkpoints.zip __MACOSX
   

7. Reproduce Figure 6 (i, ii, v, vi) for Transverse view

    python demo_fig6.py
   

Prepare the datasets

Before training the image- and projection-domain network,

YOU MUST PREPARE THE DATASET.

Train and Test the model

Train mode

1. II-Net trained at decomposition level 1 (*decomposition level 1 is equal to nstage=0)

   python main_for_img2img.py \
         --scope ct_img2img \
         --mode train \
         --num_epoch 100 \
         --batch_size 4 \
         --dir_data [PATH_OF_TRAIN_DATA_DIRECTORY] \
         --nstage 0 \
         --loss_type_img img \
         --lr_type_img residual \

2. II-Net trained at decomposition level [N] (*decomposition level [N] is equal to nstage=[N-1])

   python main_for_img2img.py \
         --scope ct_img2img \
         --mode train \
         --num_epoch 100 \
         --batch_size 4 \
         --dir_data [PATH_OF_TRAIN_DATA_DIRECTORY] \
         --nstage [N] \
         --loss_type_img img \
         --lr_type_img residual \

3. PI-Net trained at decomposition level 1 (*decomposition level 1 is equal to nstage=0)

   python main_for_prj2img.py \
         --scope ct_prj2img \
         --mode train \
         --num_epoch 100 \
         --batch_size 4 \
         --dir_data [PATH_OF_TRAIN_DATA_DIRECTORY] \
         --nstage 0 \
         --loss_type_prj img \
         --lr_type_prj consistency \

4. PI-Net trained at decomposition level [N] (*decomposition level [N] is equal to nstage=[N-1])

   python main_for_prj2img.py \
          --scope ct_prj2img \
          --mode train \
          --num_epoch 100 \
          --batch_size 4 \
          --dir_data [PATH_OF_TRAIN_DATA_DIRECTORY] \
          --nstage [N] \
          --loss_type_prj img \
          --lr_type_prj consistency \

Test mode

1. II-Net trained at decomposition level [N] (*decomposition level [N] is equal to nstage=[N-1])

python main_for_img2img.py \
      --scope ct_img2img \
      --mode test \
      --batch_size 1 \
      --dir_data [PATH_OF_TEST_DATA_DIRECTORY] \
      --nstage [N] \
      --loss_type_img img \
      --lr_type_img residual \

2. PI-Net trained at decomposition level [N] (*decomposition level [N] is equal to nstage=[N-1])

   python main_for_prj2img.py \
          --scope ct_prj2img \
          --mode test \
          --batch_size 1 \
          --dir_data [PATH_OF_TEST_DATA_DIRECTORY] \
          --nstage [N] \
          --loss_type_prj img \
          --lr_type_prj consistency \