Comparison of synthetic LGE with optimal inversion time vs. conventional LGE via representation learning: Quantification of Bias in Population Analysis

Prerequisites

• Python 3.9+

Installation

• Clone the repository and enter it:
  • git clone https://github.com/creatis-myriad/TI_selection_and_RL
  • cd TI_selection_and_RL/bin/

Create the Conda environment

• Create a conda environment from the provided file:
  • conda env create -f environment.yaml
• Activate it:
  • conda activate TI_RL_env

Data preparation

• Data come from the MYOSAIQ challenge.
• The D8 subset was not used.

TI selection

Below is the command line to get the optimal TI from the dicom files of a patient.

• Get optimal TI

  python get_TIopt.py \
      --inputDir_Dicom "/directory/of/patients' dicoms " \
      --inputDir_Roi "/directory/of/patients' segmentation masks " \
      --method "std+sat" \
      --option "std" \
      --outputDir "/output/directory"

• Visualize optimal TI
  The Figure 1 from the article was made with the command line below.

  python visu_TIopt.py \
      --load_MAG_dicoms_from_dir '/your/input/dir' 
      --load_T1_roi_file '/your/segmentation/file'
      --load_LGE_file '/your/input/file'

How to run the models

Below are the command lines to run the models:

• VAE

  python training_VAE.py \
      --inputDir '/your/input/dir' \
      --ref '/the/reference' \
      --outputDir "/your/output/dir" \
      --latent_dim 8 \
      --beta 1 \
      --epoch 80 \
      --LR 5e-4 \
      --split_data 0.15 \
      --dcm 1

• ARVAE

  python training_AR-VAE.py \
      --inputDir '/your/input/dir' \
      --ref '/the/reference' \
      --outputDir "/your/output/dir" \
      --latent_dim 8 \
      --nargs_wloss 2 10 2 20 \
      --epoch 80 \
      --LR 5e-4 \
      --split_data 0.15

Here is the command line to visualize the histogram (Figure 6) and other parameters.
It needs the output from a training, the data that you want to project in the latent space and the file with optimal TIs.

• Visualization:

  python visu_FullSeq_MAG.py \
      --input_LS_var 'latent_var_patients.pkl' \
      --load_nn_model 'save_model' \
      --inputDir '/your/input/dir' \
      --data_to_project '/your/data/to/project' \
      --ref 'reference.pkl' \
      --TIopt 'idx_TIopt_MAG_std+sat_std.pkl' \
      --ar_vae 1 \
      --fMAG 1