I understand that DenoiseImage is a spatial filtering program that removes noise assuming an underlying noise distribution (Rician or Guassian). The original paper this program is based on only discusses applications on structural scans.
Would anyone recommend using DenoiseImage instead of spatial smoothing on EPI collected during functional MRI? I’ve ran it through a few examples and the images look incredibly clean - I’m just wondering if I’m removing information from the EPI that I shouldn’t be.
Wondering if anyone has any opinions before I actually put in the legwork to see how denoising with his method impacts downstream analysis.
Title of original ref:
Adaptive Non-Local Means Denoising of MR Images With Spatially Varying Noise Levels
images (both EPI and T1w) look indeed much cleaner
this is a computationally intensive algorithm (will take a long time running this on every single EPI volume)
I did not see any analysis of how this influences tSNR or statistical power (or in context of T1w images how it interacts with VBM or surface reconstruction). If anyone knows of such analysis I would love to read it.
Just to add a note: adaptive NLM is typically used on diffusion MRI preprocessing so it seems it may make sense for EPIs. Please note it would also enhance drop-out regions, accentuating signal loss.
I worked on studying the properties of adaptive non-local means a very long time ago. There is no particular reason that this wouldn’t be beneficial for fMRI images as well given that smoothing is already a widely used procedure.
Non-local means is just smoothing on steroids and adaptive non-local means is optimal in the sense that there doesn’t exist another procedure that can do better. The main benefit of adaptive NLM is that it will not over-smooth data near boundaries i.e. it will try to minimize noise/variance with lesser bias than the usual smoothing procedures.
For fMRI in particular, if you are doing something like resting state or task BOLD correlations, then it makes sense to try to smooth only voxels together that share similar time-series to begin with. A few papers have investigated this like
That said there might be benefits to improving spatial SNR without taking temporal similarity into consideration or doing a clever combination of both. I haven’t come across anything like the latter.
I haven’t seen anyone quantify basic QC metrics that Chris pointed out.
