Checkerboard artifacts in mouse model nilearn CanICA analysis

[Ben-Abdallah]

Hello,

I’m seeking assistance with a persistent issue I’m encountering while performing ICA on mouse brain rs-fMRI data. Despite my efforts, checkerboard artifacts continue to appear around the components, and I’m unsure of the best approach to eliminate these artifacts.

For context, I’ve employed a standard preprocessing pipeline that includes slice timing correction, bias correction, motion correction, spike sorting, segmentation, normalization to the Allen Institute atlas, and initial smoothing with a 4x3x2 voxel kernel (I intentionally applied this initial smoothing to minimize the need for further smoothing during ICA), filtering in the range [0.01; 0.22].

I’ve used the CanICA algorithm from nilearn with the following parameters:

canica = CanICA(mask=mask, n_components=Ncomponent, memory="nilearn_cache", memory_level=2, verbose=0, random_state=0, detrend=False, threshold='auto', standardize=True, n_init=10, low_pass=None, high_pass=None, mask_strategy="background", n_jobs=1)

I would greatly appreciate any suggestions!

[dowdlelt]

I don’t know if this will be helpful at all, but those look like interpolation artifacts that arise from high order (or sinc/Fourier) based interpolation methods. I wouldn’t expect those to come out from a PCA/ICA step, but if these components are transformed after they are made its possible I think.

Are those values, in the black rings, very low intensity numbers? How are you viewing the images? Is there any sort of interpolation involved in producing these images?

[Ben-Abdallah]

Thank you for your prompt response. The voxel values in the black areas are approximately -0.02 when visualizing the components using tools like mrview. The images above were generated using the plot_stat_map function from the nilearn library, with the following parameters:
plot_stat_map(cur_img, bg_img=atlas, display_mode="mosaic", title="Comp %d" % (i+1), cut_coords=12, colorbar=False)
To the best of my knowledge, there is no explicit interpolation applied during the generation of these images…

[dowdlelt]

That value makes me think its interpolation related, but I struggle to see how it could have occurred. I’m not familiar with the code that was used - and even if motion correction or normilization was performed with some fourier-type inerpolation, i don’t imagine it would show up in component maps (but could be wrong). I’m at a bit of a loss.