Query about HD-BET version and running without ground-truth masks

[KeithGeorgeCiantar]

Hello Sarthak, Stefan, and the rest of the team!

I went through your paper and found the idea of an automated brain mask check very useful. I have a few questions:

• What version of HD-BET did you use? Since an updated version was released a few months ago and it was not specified in the paper.
• Would it be possible to use the tool without a ground truth set?
• Is there a way to extract a generic "quality" metric? Or is this always done by comparing with the distribution/clusters of a ground truth dataset?

Thanks in advance!

[sarthakpati]

Hi @KeithGeorgeCiantar,

Thank you for your interest in our work!

• I will leave @stefanwagner19 and @Geeks-Sid to answer the HD-BET version. I can let you know that it was not the latest version that was released a few months ago.
• Technically, yes, but. it would probably not provide you with much information since a baseline/ground-truth dataset is needed to build the clusters in its current form.
• The current implementation is based on comparing it with the distribution/cluster of a baseline/ground-truth dataset.

[KeithGeorgeCiantar]

Hi @sarthakpati,

Thank you for the quick reply!

It makes sense that you would compare with a baseline dataset, especially given the amount of variances between institutions. This way seems to make things as portable as possible.

Would it be possible to have a "default" KMeans model, similar to what was used for the paper? This way, if the user doesn't have access to a baseline dataset, a saved model (pickle or joblib) would be loaded instead.

[sarthakpati]

Hmm, that's a good point. Let me check with our data contributors' policies if this can be cleared and get back to you.