Uuummm not exactly. They are not coefficients. They are the actual regressed variables! Dice for hippocampus, Dice for whiter matter, Dice for cortex, etc.

As for the features: we have never tried to open the “black box” of the neural network… but if you do, and you find something interesting, please share!

Cheers,

/Eugenio

--

Juan Eugenio Iglesias

http://www.jeiglesias.com

From: Marco Ganzetti <marco.ganzetti@biogen.com>
Date: Monday, June 30, 2025 at 05:47
To: Freesurfer support list <freesurfer@nmr.mgh.harvard.edu>
Subject: [Freesurfer] Re: QC Synthseg

        External Email - Use Caution        

Dear Eugenio,

Thank you for sharing the reference. From what I understand, the QC scores represent regressor coefficients derived from a regression algorithm.

Is the main idea to use certain features extracted from the image to predict segmentation errors? If so, could you share which types of features you are using in SynthSeg?

Best regards,
Marco

From: Iglesias Gonzalez, Juan E. <JIGLESIASGONZALEZ@mgh.harvard.edu>
Date: Saturday, 28 June 2025 at 16:57
To: Freesurfer support list <freesurfer@nmr.mgh.harvard.edu>
Subject: [Freesurfer] Re: QC Synthseg

EXTERNAL SENDER

Dear Marco,

This QC scores represent the expected Dice of the segmentations. Of course, computing the real Dice requires reference segmentations, but this estimate is pretty good at picking up bigger segmentation mistakes, insufficient field of view, etc.

One way to use the QC scores is to reject the cases with one or more values under a threshold, e.g., 0.7.

You can find further details in this paper:   https://www.pnas.org/doi/10.1073/pnas.2216399120

Cheers,

/Eugenio

--

Juan Eugenio Iglesias

http://www.jeiglesias.com

From: Marco Ganzetti <marco.ganzetti@biogen.com>
Date: Friday, June 27, 2025 at 06:09
To: freesurfer@nmr.mgh.harvard.edu <freesurfer@nmr.mgh.harvard.edu>
Subject: [Freesurfer] QC Synthseg

        External Email - Use Caution        

I would like to know how to interpret the QC output from the SynthSeg application. What do they mean and how they are computed?

Marco