SIMPLE

SIMPLE is a simultaneous multi-plane self-supervised learning method for isotropic MRI restoration from anisotropic data.

link for arxiv paper: https://www.arxiv.org/abs/2408.13065

Installation

To use this project, use the following steps:

1. Clone the Repository:

   git clone https://github.com/TechnionComputationalMRILab/SIMPLE.git

2. Install Dependencies (via conda)

   conda env create -f environment.yml
   

Pre-Processing

You should provide a csv file that contains a table with 2 or 3 columns. Each column represents the path for MRI case (DICOM/Nifti format) in specific plane. The column name can be 'coronal'/'axial'/'sagittal'. Each row represents the cases for specific patient. The csv file should be located under SIMPLE directory.

For example:

Training

This project contains 2 models: ATME and SIMPLE.

ATME is used as a preliminary stage for creating HR MRI images. In order to train SIMPLE, you should train first 2 or 3 ATME models for the coronal, axial and sagittal planes separately and then evaluating each of them on the whole dataset.

For both models you must specify the following base flags:

--isTrain

--model_root (directory name for model outputs)

--csv_name (csv file name)

--data_format (dicom/nifti)

--global_min / --global_max (specify the grayscale range of the images, else the code calculate the absolute minimum and maximum values)

--phase (train/test)

--vol_cube_dim (the dimension of the resulted cube MRI volume - can be any value above 256). *Pay attention that different cube dimension may require different number of discriminator layers for ATME training (for example: 256 require 3 discriminator layers, 512 require 4 discriminator layers)

--calculate_dataset or --no-calculate_dataset (whether to perform pre-processing for the dataset or not. Data pre-processing must be done before the training).

*For more flags, please see 'options' directory.

• For training ATME, run 'train.py atme' command with the base flags and specify also the following flags: --plane (coronal/axial/sagittal), --TestAfterTrain (whether to perform evaluation to the ATME model immediately after its training).

  *In order to evaluate the ATME model separately you can add the flag --no-isTrain instead the flag --isTrain to train.py script or to use the test.py script (see evaluation section).

  **You can train SIMPLE model only after evaluating ATME model in each plane.

  Example:

  python train.py atme --isTrain --plane=coronal --model_root=atme_coronal_output --csv_name=<file_name>.csv --vol_cube_dim=256 --calculate_dataset 

• For training SIMPLE, run 'train.py simple' command with the base flags and specify also the flag --planes_number (specify how many planes the model is based on).

  Example:

  python train.py simple --isTrain --planes_number=2 --model_root=simple_output --csv_name=<file_name>.csv --vol_cube_dim=256 --calculate_dataset 

Evaluation

• for evaluating ATME, run 'test.py atme' command with the base flags and specify also the flag --plane (coronal/axial/sagittal).

  Example:

  python test.py atme --plane=coronal --model_root=atme_coronal_output --csv_name=<file_name>.csv --vol_cube_dim=256

• for evaluating SIMPLE, run 'test.py simple' command with the base flags and specify also the flag --planes_number (specify how many planes the model is based on).

  python test.py simple --planes_number=2 --model_root=simple_output --csv_name=<file_name>.csv --vol_cube_dim=256

Pre-Trained Model

For evaluating the model (ATME/SIMPLE) using pre-trained model, please add the flag --pre_train_G_path to specify the pre-trained generator path.

For evaluating ATME using pre-trained model, please add also the flag --pre_train_W_path to specify the pre-trained W network path.

Contact

Please contact us on [email protected]

References

• Edgardo Solano-Carrillo, Angel Bueno Rodriguez, Borja Carrillo-Perez, Yannik Steiniger, and Jannis Stoppe. Look atme: the discriminator mean entropy needs attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 787–796, 2023
• Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125–1134, 2017.