Generative Aging models for MS brain atrophy quantification.

Code for the Master Thesis: "Generation of synthetic longitudinal magnetic resonance images of subjects with multiple sclerosis: Assessment of brain atrophy and their clinical impact".

Installation

• Clone this repository:

  git clone https://github.com/anaharrismatnez/Aging_model_MS.git
  

• The clinical-super-mri repository was used to train the Super Resolution (SR) Network. If you want the repository with all submodules, clone it with the following command:

  git clone https://github.com/bryanlimy/clinical-super-mri.git
  

• It is highly recommended to create a new environment. The required packages can be found in the file requirements.txt. For the SR network see how to create the environment in clinical-super-mri.

Dataset

• Dataset should be structured as follows:

dataset/
    1_MRI_001/
        MRI_001.npy
        MRI_001_mask.npy
        info.json
        ...
    2_MRI_001/
        MRI_001.npy
        MRI_001_mask.npy
        r_MRI_001.npy
        info.json
        ...
    3_MRI_001/
        MRI_001.npy
        MRI_001_mask.npy
        r_MRI_001.npy
        info.json
        ...

• Where t_MRI_001 corresponds to each time point of patient 001, being t=1 the baseline scan, "MRI" the dataset specification, and "001" the subject identifier. As the training will be per pairs it is necessary at t > 1 to specify the follow-up scan by writing "r_" before the patient's name, meaning registered image.
• The image "MRI_001.npy" should be skull-stripped and the "MRI_001_mask.npy" file corresponds to the binary mask.
• Info.json has to contain at least the following entries about patient scan:
  • info['shape'] = [x,y,z]
  • info['delta'] = delta value between follow-up and baseline image
• Additionally info:
  • info['baseline'] = baseline acquisition date
  • info['follow-up'] = follow-up acquisition date
  • info['difference'] = info['follow-up'] - info['baseline'] (in months)
  • info['age'] = patient's age at follow-up acquisition date

• To convert .nii files into .npy, or viceversa use the following commands:

  python utils/nii_2_npy.py -source dataset -output_dir data_npy
  

  python utils/npy_2_nii.py -source data_npy -data_path dataset 
  optional arguments:
  -resize
        True if .npy files are obtained from cGAN network output.
  
  

  npy_2_nii.py automatically saves the niftis files at folder: data_npy_niftis.

Models

• cGAN + SR. Training and inference pipelines: training_cGAN_SR.sh and inference_cGAN_SR.sh
• cLDM. Training pipeline: training_cLDM.sh.

Monitoring and Visualization

• cGAN and cLDM monitor training through wandb via anonymus mode.
• clinical-super-mri repository monitors the training via Tensorboard (check clinical-super-mri).