[Insight-developers] itkOptMattesMutualInformationImageToImageMetric.txx question

[Luis Ibanez]

Hi Andrei,


Question 1:
===========

The main difference between the itkOptMattesMutualInformation class in
Code/Review, and the itkMattesMutualInformation in Code/Algorithms is
that the itkOpt one is mutlti-threaded. There are also many small 
differences in their implementation, the most relevant of which is that
the itkOpt also give you some options for dealing with lower memory
requirements.



Question 2:
===========

About comparing the ITK implementation against Matlab.


Here is the standard rant about the "method comparison" dissease
that plages our field:


       Such an idea just doesn't make sense in real practice.


Why?

     Because the performance of the metrics (and in general any
     image processing method) is heavily dependent on the parameters
     that you use in them, as well as the input that you use to feed
     them.

     The only serious claim that could be made about comparing
     performance should include the details of:

         a) Full set of parameters used
             - Number of samples
             - Number of bins
             - Degree of Parzen windowing

         b) Build settings
             - Compiler options
             - Platform
             - Specific processor (number of threads/number of cores)

         c) Full description of the data
             - Image modality
             - Image dimensions



The recurring question of "which method is better/faster" is a cancer
that metastasized our field, and its source is the lack of scientific
education in those who review and publish papers in Journals that do
not require verification of reproducibility. (e.g. almost all journals
that *pretend* to be scientific publications)


The quest for the "best" method is as senseless as the quest for the
best food. I can always claim that my recipy for making packaques is
better than anybody else's.


The question about "method comparison", is what Journal reviewers, and
Ph.D. dissertation jurys bring up when they run out of smart questions
to ask. It generally indicates a lack of hands-on experience on the
practice of image processing.



That being said,

Serious performance comparisons between two different implementations
of image analysis methods are more than welcome to the Insight Journal.
(www.insight-journal.org). They should include, at a minimum, the
information listed in (a), (b) and (c), and of course they should
include all the source code, data and parameters that will make possible
for readers to reproduce the experiment.



      Regards,



           Luis



---------------------------
Andrei Danilchenko wrote:
> One more question about itkOptMattesMutualInformationImageToImageMetric.txx.
> In comparison with itkMattesMutualInformationImageToImageMetric.txx, is it
> optimized for smaller computation time or the difference between them is the
> structure of the class? Do you know how much faster ITK version of MattesMI
> performs in comparison with the corresponding code in Matlab?
> 
> Thanks,
> Andrei.
> 
>