search

khanlab / hippunfold

BIDS App for Hippunfold (automated hippocampal unfolding and subfield segmentation)
https://hippunfold.readthedocs.io
MIT License
49 stars 13 forks source link

Question: QC/unetf3d_dice.tsv #259

Closed andreifoldes closed 1 year ago

andreifoldes commented 1 year ago

Greetings,

I would like to ask what the _unetf3ddice.tsv files mean and how to interpret them?

jordandekraker commented 1 year ago

Most of the time, the UNet step that does tissue segmentation (eg. white vs grey matter) performs very well, but sometimes it fails catastrophically (eg. will be missing a large chunk of hippocampus, or sometimes it thinks the collateral sulcus is hippocampus). This is rare (approx 1%).

f3d is a simply deformable registration to a reference atlas that has a corresponding hippocampal label. This is a very coarse, non-detailed way to obtain a hippocampal segmentation, but by checking that there is at least good (eg. >0.7) Dice overlap with the UNet result, we can be assured that the UNet did not catastrophically fail. The Dice overlap will never be perfect (ie. near 1.0) since the UNet result SHOULD be more detailed, and so as long as this value is above 0.7 the result is likely good and there's no need to interpret it further.

I hope that helps!