Dear colleagues,
I'm interested in using BrainVISA's Morpohologist pipeline to study the hemispheric asymmetry in the length of the Sylvian Fissure, which displays a population-wide asymmetry towards the left hemisphere. As a test, I first entered the original T1-anatomy images of participants as well as their left-right mirrored versions into the pipeline, expecting the reconstructed Sylvian fissure to reverse as well. However, I found that the Sylvian fissure length of the original and left-right flipped brains had the same - usual leftward - asymmetry, albeit that the degree of asymmetry was weaker for the artificially flipped brains. I visually checked all reconstructed fissures to make sure this was not due to differences in the labelling of segments between the original and flipped brain. In doing so, I noticed that the overall shape of the Sylvian Fissure likewise did show the dramatic reversal I expected in the left-right mirrored brains. I believe this may be indicative of bias, which I think most likely originates from the asymmetry inherent to the BrainVisa sulcus atlas. I was wondering if simply using a symmetrical version of the default sulcus atlas would fix the issue and, if so, how can I use such an atlas within the morphologist pipeline? If not, what additional/other steps should I take?
Morphologist - Symmetric atlas to reduce bias?
-
riviere
- Site Admin
- Posts: 1361
- Joined: Tue Jan 06, 2004 12:21 pm
- Location: CEA NeuroSpin, Saint Aubin, France
- Contact:
Re: Morphologist - Symmetric atlas to reduce bias?
Hi,
In Morphologist, the segmentation part is not lateralized, and thus will process a reversed image the way you expect. If I understand this is what you see, the segmented image and extracted folds show the reversed asymmetry you were expecting.
Now the sulci recognition is using an asymmetric atlas, you are right, and it will probably not assign sulci identification in a perfectly correct way to a symmetrized brain. The posterior part of the sylvian fissure is especially asymmetric, and in our altas we even have an additional sulcus on the left side that we do not (or too rarely) observe on the right hemisphere (the S.GSM sulcus), which of course has an influence on the surrounding sulci identification. So I guess that the sulcal labelings on a symmetrized brain are not completely correct (perhaps a part of the FCL is dropped or added), and tend to "pull" things toward the atlas.
What we do more frequently when we deal with such studies is to process the brain in the regular way (non-symmetrized), extract sucli, then flip the brain after sulci identification. Of course there is still the possibility of a bias due to the sulcal identification model which tends to impose its asymmetry to your data.
Otherwise symmetrizing the model is not very easy since it includes both images and probabilistic priors.
In Morphologist, the segmentation part is not lateralized, and thus will process a reversed image the way you expect. If I understand this is what you see, the segmented image and extracted folds show the reversed asymmetry you were expecting.
Now the sulci recognition is using an asymmetric atlas, you are right, and it will probably not assign sulci identification in a perfectly correct way to a symmetrized brain. The posterior part of the sylvian fissure is especially asymmetric, and in our altas we even have an additional sulcus on the left side that we do not (or too rarely) observe on the right hemisphere (the S.GSM sulcus), which of course has an influence on the surrounding sulci identification. So I guess that the sulcal labelings on a symmetrized brain are not completely correct (perhaps a part of the FCL is dropped or added), and tend to "pull" things toward the atlas.
What we do more frequently when we deal with such studies is to process the brain in the regular way (non-symmetrized), extract sucli, then flip the brain after sulci identification. Of course there is still the possibility of a bias due to the sulcal identification model which tends to impose its asymmetry to your data.
Otherwise symmetrizing the model is not very easy since it includes both images and probabilistic priors.