Deconvolution vs. AI 'Deblurring' for microscopy

Does anyone here have experience using or comparing general deconvolution methods with ‘AI’-based ‘de-blurring’ methods for microscopy data? I have seen some papers in the literature touching on this.

I use the term ‘general deconvolution’ to mean mathematical methods that aim to recover the un-convolved image from a knowledge of the input image +/- PSF(s) only (this includes what basic priors can logically be known from them like non-negativity and any theoretical maximum intensity level in the solution).

Personally, I don’t consider ‘AI’ (deep learning) methods that perform de-blurring to be ‘deconvolution’ because of their imputational nature attempting to fit extraneous data to a solution regardless of how plausible and wonderful the results may look. Am I alone in this view? Unfortunately (from my perspective) there are increasing numbers of peer-reviewed papers being published that permit the ‘D’-word to be used in conjunction with AI imputation methods.

Do any of you have personal experience of highly convincing yet hallucinatory results using such methods? What would be your standard for peer reviewing papers claiming to have derived new data using them (like super-resolution)? For example, how would you view it if someone trained an AI de-blurrer on EM ultrastructure and applied it to ‘deconvolve’ a diffraction limited light microscope image of a semithin or ultrathin section stained with tol blue so they could ‘resolve the mitochondrial cristae’ in the LM image?

If there are already good discussion / review / guideline resources on this subject, please point me to them.
I will re-post on imagesc in case others in the know don’t frequent this corner.



Just to illustrate, this is the kind of nonsense I am hoping awareness of this issue will avoid (although it is at the extreme end, the slippery slope does beckon):

The discussion for this topic is mostly on the imagesc version here: