Which atlas to use?

[mschoettner]

@illdopejake how do you best decide what atlas to use that you base your ROIs on? Apart from the papers you linked to in your notebook on machine learning, is there anything in particular that I should consider?

[illdopejake]

Great question! Yes, I think there are quite a few things to consider. The top three considerations are probably as follows:

1. What resolution do you want? In other words, how many regions/features? As the number of regions increases, you get potentially more information, but you also may increase the complexity of your model (in ML) or number of comparisons (in traditional stats).

2. What modality should the atlas be based off of? There are many atlases out there, and they are created based on a number of modalities, including histology, anatomical structure, functional connectivity, multiple modalities, etc. The real question you want to ask here is, what do you want your regional borders to mean? Should they be regions that are more functionally connected to one another? Regions that share similar cytological characteristics? Or do you just want random parcellations with no reference to brain structure or function?

3. Should I use an already created atlas, or base one off of my own dataset? Again, there are many atlases to choose from that have been created by other groups using a range of modalities (see below). There are many reasons to choose one of these atlases, the most prominent being that your dataset is not influencing your regions (might be important for generalizability). However, you may wish to create your own parcellation from your dataset, to capture the best capture the unique sources of variation in your own data. If you want to choose an atlas, you can source on directly from the nilearn datasets or one from this list here. If you would like to create your own parcellation from your data, I would point you to the Nilearn User Guide, particularly sections 3.3, 3.4 and 3.5.

Other resources
This paper is a great resource for learning about the thought process that goes into creating different brain atlases. This is a good one as well.

I would also highly recommend giving a read through this excellent paper from Gaël Varoquaux's group, which considers how different choices made in a standard ML pipeline can influence outcomes. In this paper, there is a comparison of different atlases.

[mschoettner]

Thanks a lot for your thorough answer! That should definitely help me in making a decision. I will have a look at the papers as well!