Selecting timeseries from multiple brain atlases



Some fmri data challenges give multiple different timeseries from different atlases as functional mri features. I am wondering why this is so and how to choose an atlas and its timeseries from the many given ? Is there a standard approach that can be followed and built upon ?


Parcellating a multivariate spatially diffuse entity (such as the brain or planet earth) into a finite set of non-overlapping regions is IMHO an ill posed question. Depending on which feature of data you focus on (connectivity? macro scale anatomy?) you will get equally well suited parcellations. Same goes with the number of parcells (most regions can be subdivided if you redefine the problem). This is why there is no (and I don’t think ever will be) one true ultimate parcellation that works well for all problems.

In research practice it’s worth checking if your results hold for a range of different parcellations.


@ChrisGorgolewski i see. ok, thanks. So, i understand this means that during training, i can select any one of these atlases and confounds and try to classify the dataset. Then repeat training iteratively with other atlases and select the one that gives the highest accuracy. Is that right ?


Yes in context of machine learning (where interpretability matters less) you can think of a parcellation as feature extraction strategy. Thus you can use cross validation to pick the parcellation that yields the best predictions.