In short, perform these procedures separately to get the confounding regressors of interest from them (i.e. the top 5 components from aCompCor and the timeseries of the ICA components identified as noise from ICA-AROMA). Then regress these confounders out using something like nilearn’s clean_img function.
EDIT: The top consideration that comes to mind is that ICA-AROMA is, in my experience, not good at identifying multiband artifacts as noise. FYI, these typically manifest as parallel lines seen in sagittal or coronal views:
Therefore, if you dataset is small enough, you might want to just manually go over the ICA components in a viewer like fsleyes and check which ones ICA-AROMA missed. If your dataset is huge, consider the return-on-investment that’s offered by ICA-FIX instead.
We developed a regression-based technique, MARSS, that successfully removes an artifact manifesting as elevated signal correlations in simultaneously acquired slices. The removal method is run directly on unprocessed data (i.e., before any spatial normalization or SDC). We also showed that ICA-FIX does not fully mitigate the same artifact, so it may be advantageous to use MARSS prior to ICA-FIX for optimal denoising. The source code is here if interested (we have MATLAB and Python implementations): GitHub - CNaP-Lab/MARSS: Multiband Artifact Regression in Simultaneous Slices (MARSS)