Evaluate and support MEGNet

[mscheltienne]

Found this randomly: https://www.sciencedirect.com/science/article/pii/S1053811921006777
Associated code: https://github.com/DeepLearningForPrecisionHealthLab/MegNET_2020

Might be interesting (I haven't read the paper yet).

[adam2392]

This looks cool!

[jacobf18]

Looks great. They use tensorflow for their model. Should be easy to convert to ONNX.

[colehank]

Hi! Dear MNE Developers,

I have integrated MEGnet into my MEG preprocessing workflow using data from a 256-channel CTF system. The artifact recognition hit rate is quite good, but on average, about one or two artifacts are missed per run. The artifacts identified so far include only blinks, eye movements, and heartbeats, as MEGnet was trained on labels for these three types of artifacts. To enhance the functionality, it may be necessary to retrain the model using more subjects, additional MEG systems, and with a broader range of artifact labels in order to achieve results comparable to MNE-ICAlabel. This is my experience using it, hopefully useful! :)

The attached figure displays the ICA components from one run, where components with red titles are those identified as artifacts by MEGnet.

Best regards,
Guohao

[mscheltienne]

Looks good, it would be awesome if you could take the time to create a PR to mne-icalabel to include MEGNet in the available algo. It would have to be a submodue in mne_icalabel/megnet defined similarly as iclabel in mne_icalabel/iclabel.

[colehank]

Sure! It would be awesome to contribute to MNE. However, I haven't been deeply involved in the development of the toolkit before. I'll take some time to study the structure of mne_icalabel/iclabel and see if I can work on adding mne_icalabel/megnet.

[mscheltienne]

Great, we will definitely give you all the help and advice possible!