Interpreting ICA-Aroma artifacts

Example of ICA-AROMA components for resting data from a GE scanner, spiral sequence


Occasionally I see ICA components where the spatial pattern is concentrated in predominantly in an axial slice (e.g. #45, #46, #47, #49, #50, #51, #52). These artifacts are not consistently classified as noise, they are occasionally designated as signal as well (e.g. #49)

  • Do people regularly see such artifacts? Is this a known pattern?
    They don’t seem to match any of the canonical artifact patterns, for instance those described in Griffanti et. al. 2017.

Hi @mnarayan, as you are suggesting since the beginning I suspect that ICA-AROMA wasn’t trained on data acquired with spiral sequences.

Moreover, ICA-AROMA was envisioned to remove head-motion related artifacts and thus, it is very possible these artifacts you are seeing come from some other source (or are artificially introduced by the much harder reconstruction of spiral sequences) - in other words, ICA-AROMA wasn’t trained to classify those components as noise.

If the artifact is shown in all or most of your data, there is room for you to work with the technicians and identify the source of this problem.

Does this cover what you were asking?


After visual inspection of the ICA-AROMA outputs, is there any way to specify by hand which to classify as noise vs signal?

Thanks, yes it does if you are confirming that you don’t notice these patterns on EPI datasets? I don’t think I have seen enough of them from different scanners, multi-band and non-multiband sequences to be able to come to this conclusion.

I don’t know about by hand. One could certainly write some code to identify sudden discontinuous changes - 2-D edge detection within 3-D spatial maps to identify these I think.

I haven’t seen that particular artifact before if that counts.

I get those too, from a GE scanner. I’ll have to update to see what our sequence is.

1 Like

update: we are using a gradient echo sequence, TR=2 seconds, TE=80 milliseconds, sequential slicing bottom up.