Dear Renzo

The components from PCA are simply sorted in decending order by percentage variance explained in white matter or CSF, so there is usually no specific interpretation expected of each individual component. Jointly, they are expected to represent a basis characterizing non-neural sources of BOLD signal variability. The optimal number of components to retain depends on each specific dataset, the "Chai et al .2012. Anticorrelations in resting state networks without global signal regression" paper presents analyses justifying that a relatively small set of components suffice to produce appropriate denoising, while the "Morfini et al. 2023. Functional connectivity MRI quality control procedures in CONN" discusses this choice in the broader scenario of quality control (where you can increase or decrease the number of PCA components included in your denoising procedure to optimize the quality of the resulting timeseries). In terms of the CONN gui, 16 is the (maximum) number of PCA components that are estimated during the Setup step, and later in the Denoising tab you can choose how many PCA components you may want to include as part of denoising from each of the White matter and CSF areas (by default 5 components each). Increasing the number 16 in the Setup step will not have any consequence, other than simply allowing you to later choose a higher number of components later during denoising if you wish to do so, and increasing the default number 5 during the Denoising step will typically result in a more conservative denoising procedure (and decreasing that number will result in a more liberal denoising procedure), which you may wish to use to adapt/optimize the choice of denoising parameters for your own dataset.

Hope this helps

Alfonso

Originally posted by Renzo Torrecuso:

Dear Alfonso or CONN experts,

Could you please help in these 2 questions?

1- If I got it right,  CONN bases its PCA for white matter on Behzadi et al 2007, but the latter proposes the "broken stick" method and does not specify a specific number of PCA components.

 Could you please explain the motivation to use 16 components in CONN?

2- From what I understood, the last components (from 6 to 16) might reflect physiological noise (e.g. aliased cardiac or respiratory signals) or scanner-related artifacts not shared across regions, and these subtle patterns can reflect microvascular or motion-induced effects often missed by dominant components.

Would you happen to know if there is any explanation on what the 6th, 7th and so on until the 16th might be explaining? 

 

Thank you very much.

All best

Renzo