Discussion about the smoothness of the GM mask based on the AMU15 atlas

[jcohenadad]

While working on re-generating the GM to be more truthful to the AMU15 atlas (see notably #41, #45), I stumbled upon some considerations.

Here is the original comment from @vcallot's #32 (comment)

The AMU15_GM had smoothly varying probability – A threshold at 0.5 would lead to a conservative mask, avoiding PVE in WM
The PAM50_GM presents very sharp changes (from 0.01 to 0.7 changing just one pixel) – A threshold at 0.5 leads to  (from my point of view) too large GM, especially in the Gray matter commissure for instance. The probability at this location is around 0.9, and the commissure is 2 pixels large, which is not very representative. I did not really pay attention to that previsously because we erode our mask before quantification to be on the safe side with regards to PVE contamination and we don’t use the maximum likelyhood or average option.
Also, using a thresholf of 0.5, if we compute the ratio of GM/SC areas, we end up with a ratio of 20% with PAM50, whereas it was 15% with AMU15/MNI_poly_AMU. So this may have an impact for quantification, fMRI studies, etc ..

These are all excellent points, however there are two things that worry me.

1. The probabilistic values from the original AMU15 GM atlas might be impacted by the registration parameters. That registration, which I assume was solely based on the SC mask to co-register the 15 images together (to be confirmed), was likely non-linear (eg: SyN or BSpineSyN algorithm). If there were strong deformations inside the spinal cord, these could lead to an under-estimation of the probability values for each GM and WM class. In other terms, we would expect boundaries to be less 'soft'. In order to address this, we would need access to the original data and masks (SC and GM), and the cord that was used to generate the AMU15 atlas.

2. In cases where users register data using the T2*w image, the probabilities of the GM mask should be increased. In scenario where users have a T2*w image and use it to register their data to the PAM50 template, the concept of a probabilistic GM mask becomes irrelevant. What becomes relevant in this case is the encoding of the partial volume. This points highlights that, depending on the usage of the PAM50 GM mask, users might either want probabilistic values (reflecting inter-individual differences), or partial volume values. The current PAM50 GM mask is closer to a partial volume mask, while the AMU15 is closer to a probabilistic mask. We need to decide what should be encoded in the PAM50 GM mask.

[vcallot]

@jcohenadad and SCT team, please see : DOI: 10.1007/s10334-013-0403-6 for the detailed construction of the AMU 15 T2* template + GM/WM probabilistic atlases, and their definitions ("probabilistic" rather refers to the probability to belong to GM or WM tissues than inter-individual differences). Hope this helps.

[vcallot]

2013_MAGMA_Taso_SCAtlas.pdf

[jcohenadad]

Thank you Virginie, it does help. The registration method was in fact linear (see p3):

This addresses the concern in point (1), ie: no non-linear deformation is to be expected inside the spinal cord.

From #50 (comment)

"probabilistic" rather refers to the probability to belong to GM or WM tissues than inter-individual differences

I partly disagree with that statement. While the 4-DOF (ie: rotation, Tx, Ty, scaling) of the "probabilistic" template does account for cord size differences across individuals compared to the 3-DOF version, it does not account for non-linear differences in terms of GM shapes and location across individuals. The traditional approach for creating a template is to use non-linear registration, to precisely address subtle differences of sub-structure across individuals. Hence, I do believe that the 4-DOF atlas does include inter-participant differences.

Which brings me to my second concern, point (2): If users register images using a WM mask (eg: originating from a high resolution T2*w scan), where WM/GM boundaries are clearly defined, then mapping an overly smooth WM/GM mask would defeat the purpose of working with a high resolution scan that has sufficient contrast to distinguish GM from WM, and in which case only PVE should be encoded (not inter-participant variability).

[jcohenadad]

After discussions with various collaborators (incl. folks in Zurich), I'm leaning towards re-generating the GM mask based on non-linear registration, in an iterative manner (as done by ANTs template generation pipeline), such that the output GM mask would mostly encode partial volume information.

Re-generating it using a unified pipeline would also be required if we wish to combine the AMU data with the Zurich lumbar data to extend towards the lumbar cord #45.