search

fabianhielscher / AUTOMAP

This is an implementation of AUTOMAP which reconstructs MRI-images with undersampled k-space data
7 stars 3 forks source link

readme

AUTOMAP

This is an implementation of AUTOMAP which reconstructs MRI-images with undersampled k-space data Help1

Setup

  1. Clone the repository
  2. Create a virtual environment with python 3.6 (3.7 does not work)
  3. Install the following packages:
    • numpy 1.14.5
    • scikit-learn (sklearn) 0.21.3
    • tensorflow 1.7.1 (tensorflow 2.0 does not work)
    • pillow 1.1.7

Manual

  1. preprocess_images.py\ This module preprocesses raw images in the images_raw directory. Processed images are stored in the root directory.

  2. neural_network_train_and_save.py\ This module trains the neural network, undersamples the data with undersampling masks in the "pattern"-directory, creates a directory named "saved_models" and saves the network in this directory.

  3. neural_network_use_trained_model.py\ This module uses a trained module and saves the output images from given inputs. The reconstructed images are stored in folder that is created and named "recon". A meta-file is created during the training process located in "saved_models". The name of this meta-file needs to be copied to the variable "saved_model_meta_info". This information needs to be set to load the correct model.

Help1 Help1

Options:\ Change the target width and height from n=64 to the desired resolution. Higher resolutions might cause memory issues. Use different training images by replacing the folders in images_raw.\ e.g. images_raw/custom_folder/custom_images (images must be in folder inside images_raw).

Hyperparameter tuning

Help1

Reconstruction examples compared to compressed sensing

Experiment 1

Help1

Experiment 2

Help1