MEGNET

This repository is a fork of the code listed below in the original code reference. This repository adds an automated processing wrapper and python package installation around the original codebase. The current codebase utilizes mne python to preprocess the data, generate the infomax ICA components (n=20), circular topography maps, and timeseries outputs. The architecture of neural net has been preserved, however, the weights have been reset to uniform distribution and retrained using repository data from MEGIN, CTF, 4D, and KIT systems.

Install (pytorch based version)

Tested compatability: (mne 1.10/1.11/1.12 & python 3.11/3.12/3.13/3.14)

conda create -n megnet 'mne>1.10,<1.13' 'python>3.10'
conda activate megnet
pip install MEGnet-neuro
megnet_init  #Download model weights from hugging face

Conda package

The conda-forge v1 recipe and local build instructions are in conda/. Until the recipe is accepted and published by conda-forge, build it locally with rattler-build. After publication, install the package with:

conda create -n megnet -c conda-forge megnet-neuro
conda activate megnet
megnet_init  # Download model weights from Hugging Face

Install (tensorflow based version)

Tested compatability: (mne 1.10/1.11/1.12 & python 3.10/3.11/3.12/3.13) (py restricted by tensorflow builds)

conda create -n megnet 'mne>1.10,<1.13' 'python<3.14'
conda activate megnet
pip install 'MEGnet-neuro-tf'
megnet_init #Download model weights from hugging face

Original Code Repository

https://github.com/DeepLearningForPrecisionHealthLab/MegNET_2020
Manuscript available: https://pubmed.ncbi.nlm.nih.gov/34274419/
DOI: https://doi.org/10.1016/j.neuroimage.2021.118402
PMID: 34274419