BrainVISA

Hi,
I am trying to understand the variability in results of Morphology 2015 pipeline.
I have processed the same MRI scan a few times using exactly the same parameters and options.
Then, I opened the produced cortical fold graphs in the viewer.
Well, after a very brief look I have seen that the graphs vary. Even ignoring the differences in labels, there are some areas where the graphs look very different for every run (please see the marked area on the attached image). So now I am trying to understand what happens.

• Is this variability in results expected and why (i.e. what part of the pipeline causes different cortical segments to be extracted at every run of the pipeline)?
• Could this variability be minimized (or even eliminated) by carefully choosing options and parameters for the pipeline? I don't mind having different labels, but is it possible to have exactly the same structures for the unlabeled graph every time, i.e. extracting the same nodes at the same location?

Thanks a lot,
Ruslana.

Hi Ruslana,

Yes there is some variability in the processings. Many of the algorithms in Morphologist are using stochastic optimization to overcome (partly) local minima problems, so different runs do not produce exactly the same results.
On the segmentations variations are observed generally in low contrast or noisy areas where image quality is problematic, or on shallow folds (which may or may not pass a threshold with slight segmentation variations).
There may also be variations on the number of pieces a fold is split into, also when around the threshold of a split criterion.
We must admit that we have not made a thorough study of the overall variability.
Otherwise as you can see major, deep and well marked folds are always extracted, in a rather consistent way, so the problem is normally limited to "borderline" regions (most frequently, from my experience: temporal pole, basal frontal and temporal regions, calcarine sulcus, sometimes occipital pole - which are regions causing more problems to all segmentation algorithms, actually).
It is possible to disable this variability from one run to another, by fixing the random seed used in stochastic processes (there is a parameter for that in Morphologist), however even if it will produce the same result in all runs, it will not ensure those results are the best, and it will just "hide" the actual variability. Normally this fix_random_seed option is only used for regression tests that we run internally to compare results to reference reference ones.

Denis

Denis,
Thanks a lot for clarification. It helps a lot!
Regards,
Ruslana.