SLED

Self-Labelled Encoder-Decoder (SLED) for myelin water imaging data analysis

How SLED works:

• Data: multi-echo myelin water imaging (MWI) data
• Encoder: a series of neural networks to estimate latent parameters such as T₂ or T₂^* times and amplitudes
• Decoder: a typical 3-pool model (myelin, axonal, and free water pools)
• Training: exclusively trained for each dataset which is self-labelled

File structure in this repo

• configs: configuration files for all the fitting initialization parameters
• data: imaging data in nii format are stored here by default (change the path in config file if the data are located elsewhere)
• models: trained model (the best epoch) is saved here by default
• results: parameter maps are saved here by default
• src: source code is placed here
  • models: building blocks for the SLED model and the code for model training
  • utility: utility functions for image processing and customized tensorflow losses
  • main.py: entry point to the SLED fitting

Run SLED fitting

• Clone this repo to your local computer

  git clone https://github.com/hanwencat/SLED.git
  

• Nevigate to the root of this repo (should be on the exp_3pool branch by default, not the master branch)
• Use conda to create a virtual environment for the SLED fitting

  conda env create -n sled -f environment.yml
  conda activate sled
  

• Modify the config file as needed (e.g. the data_path in the configs/defaults.yml)
• Run the src/main.py to fit

  python src/main.py
  

• The model training process is logged in logs/training.log
• Fitted maps are saved in the result/ folder
  • mwf.nii.gz: the fitted MWF map
  • t2s.nii.gz: the fitted t2 times for the 3 pools (4D data)
  • amps.nii.gz: the fitted amplitudes for the 3 pools (4D data)
• The fitting configuration (everyting in the config file) are stored as nifty header extension (NIFTI_ECODE_COMMENT) for all fitted maps (https://nipy.org/nibabel/devel/biaps/biap_0003.html)

Reference

Please cite this paper https://doi.org/10.1016/j.neuroimage.2022.119717