Home AutoQuant General Discussions

Measured PSF tool?

I am processing Light Sheet datasets, which requires measured PSF for 3D deconvolution.
Where is the PSF extractor/distiller tool, if there is one?
I have the beads dataset, just not sure how to go about averaging PSF from multiple beads.I guess I could just pick a single bead and crop, but averaging multiple PSFs may be better.
Huygens has the PSF distiller, so I expect Autoquant has something similar, just cannot find it.



  • Hi Stanvitha, Many thanks for your question. 

    At present, a PSF averaging tool is not available in AutoQuant.  We have done a lot of work recently developing the AutoQuant module for Image-Pro, and many development plans going forward including the addition of an averaging tool. 
    For now, you will need to crop a single representative bead stack. 
  • Thanks, I will give it a try. 
    Are there any strict requirements for the PSF in terms of size (number of pixels pixels -  power of 2? - Also, I assume the PSF does not have the quadrants inverted, like e.g. the COSM or XCOSM software used to do).
    How precisely does the PSF need to be centered in the cropped mini-stack?
  • Hi,
    There are no strict requirements on the PSF size nor on how centered it is (though I recommend having it roughly centered to minimize the potential for information loss). AutoQuant will handle the necessary pre-processing internally, including de-noising, resizing, centering, and quadrant inversion.

  • I'd be very careful using the Autoquant measured PSF algorithm.  In their recent Autoquant white paper the measured PSF performed very poorly.   See table 9 here https://mediacy.com/wp-content/uploads/2023/03/An-evaluation-of-the-efficacy-of-adaptive-and-fixed-PSF-deconvolution-using-the-AutoQuant-module-of-Image-Pro-version-11.pdf

    What was surprising about this result is that the author of the white paper, mis-labeled a theoretical PSF as 'measured'.  So in fact what happened was a theoretical PSF, which had undergone the Autoquant measured PSF pre-processing, performed poorly.  A theoretical PSF should not be changed much at all with preprocessing (because it is already noiseless and centered).  

    This indicates there could be a bug in the measured PSF pre-processing routine, or alternatively there could have been a meta data issue or mistake entering parameters.  Or there could have been a mistake in the PSF the 3rd party group (EPFL) provided.   However I would not be confident in the measured PSF preprocessing routines in Autoquant, until they clarify why the measured PSF algorithm performed so poorly in the white paper. 
  • That is quite interesting.  
    I will report back how it goes - my first attempts with the measured PSF (a cropped stack containing a single bead image) reliably causes the Autoquant X.3 to crash. I will test if unselecting "GPU processing"  will make a difference.
    our facility also has Amira with the deconvolution option, I will do a test run on that as well and compare the outcome. 
  • Stanvitha, I'm sorry to hear you ran into a crash while using a measured PSF. I'll contact you for details to help troubleshoot.
  • Hi Andy,
    I am running X3.1.3
    Green fluorescent beads, 1920 pixels square, 243 z slices  (16bit data; 1.7GB file size)  .
    Enable GPU processing is on.
    Subvoluming (9 subvolumes) performed OK, after it starts processing the first subvolume, it crashes after about 1 min.
  • the same dataset with CPU processing was processed OK. 
  • Hi stanvitha

    I took a close look at the theoretical PSF that was provided by EPFL, and used in the Autoquant white paper, table 9. I now am suspicious that the PSF provided with the dataset is in fact incorrect, and the differences in results are not because of anything to do with preprocessing the PSF in Autoquant. 

    Below screenshot shows the EPFL PSF on the left, and a theoretical PSF generated with the parameters they provide on the right, note the difference in cone angles.  I double checked this a couple of different ways, and am planning to follow up with EPFL to confirm and figure what happened, as this dataset is used in lots of comparisons. 

    I still would be interested to hear about any experiments you do deconvolving images using measured PSFs with different programs.  As subtle differences in PSFs can make a big difference in Deconvolution results and the preprocessing is different between different implementations. 


Sign In or Register to comment.