Jeff and others,


> I would rather see real beads.
>

Sorry, but I don't agree on this point. Using simulated data has the
advantage that you know exactly what is your input to the algorithm,
thus enables you to judge how the algorithm treats different kinds of
aberrations. E.g. you create an "ideal" object and can apply all kinds
of PSFs (also spacially varying ones), noise, lamp/laser jitter, ...
separately by software. If you take "real" data, you always get the same
set of aberrations that the microscope(s) you use create(s).
I think the data Andres shows is a step in the right direction. Still I
have two comments:

- Beads as test objects don't show everything. As MLE and similar
algorithms are nonlinear processes, the quality of deconvolution should
depend on what kind of object one looks on. Maybe over time the
microscopy community can agree on what test objects are suitable (dots,
rods, filaments, fractals, E-M images of cells, ...?) In the end one
would have a generally accepted set of test data that everyone, who
wants to test a deconvolution algorithm, can use.
- It is hard to judge by eye how good a result is. What does it mean, if
a bead looks a bit smaller after deconvolution? Does the algorithm just
"eat" the borders (overreconstruct) or is this a sound regain of
information? I would be much more confident to judge a result if I had a
generally accepted figure of merit that tells me how good it is. Again,
there will probably be several suitable figures of merit that all have
their pros and cons. So one would like to apply them all to get a
general impression. And again, one needs to use simulated data to be
able to compare the deconvolution result to a real original.

To give you an impression of what I mean I would like to draw your
attention to the work of a collaborator of ours, Rainer Heintzmann.
http://www.aphys.uni-heidelberg.de/AG_Cremer/rheintz/Dissertation/node55.html
I have not been involved in this work, so this is no self advertisement.
I just found myself in the situation of having to choose a deconvolution
algorithm for our lab and having to rely on in a way subjective
criteria.

This mail is also a response to mails that Michael C. Wussow and Beat
Ludin posted to several people off the list. Michael, Beat, would you
like to post your mails to the list, again?

Joachim


--
------------------------------------------------------------------
Joachim Walter, Dipl. Phys.
Institut für Anthropologie und Humangenetik der LMU München
AG Cremer
Richard-Wagner-Straße 10/I
D-80333 München               Tel. +49 - 89-2180-6713
Germany                       Fax. +49 - 89-2180-6719
------------------------------------------------------------------