Masks drawn based on tSNR image instead of mean

[MerveKaptan]

Hello @jcohenadad and @rohanbanerjee

One of the sites that would like to share their data has drawn their masks based on tSNR image instead of the mean image. Although the match between the mask and the mean image seems to be somewhat fine, there are some slices where this seems to be an issue. Below, you can see a case where the manually drawn mask is clearly on CSF

Obviously, we would need to trust our ground truth as we were talking about it in this issue, but in this case, there is no ambiguity (the mean image and mask clearly do not match).

So, I was thinking what you would recommend doing wrt this dataset. Should I look through all subjects x slices and try to identify how frequent this mismatch is?

Merve

[jcohenadad]

One of the sites that would like to share their data has drawn their masks based on tSNR image instead of the mean image. Although the match between the mask and the mean image seems to be somewhat fine, there are some slices where this seems to be an issue. Below, you can see a case where the manually drawn mask is clearly on CSF

I don't think the issue is specifically caused by drawing the masks on tSNR images. I could imagine similar issues with masks created on the magnitude GRE-EPI data.

So, I was thinking what you would recommend doing wrt this dataset. Should I look through all subjects x slices and try to identify how frequent this mismatch is?

Coming up with a systematic and well-documented approach to generating GT data is important. Information about who did/reviewed the GT should be stored in the JSON derivatives. @rohanbanerjee pls provide the necessary info based on our current practice in the lab (with links to our internal doc, etc.)