[Insight-users] Normalisation of images necessary?

[Luis Ibanez]

Hi Jef,

Normalization of the image intensities is not required for
the Image Metrics:

 > MutualInformationHistogramImageToImageMetric
 > NormalizedMutualInformationHistogramImageToImageMetric

However, what is *VERY* important is to make sure that you use
the range of intensities that is relevant to the anatomical
structures that you care to register.

In other words, your image will have section of the dynamic
range of intensities that are not contributing (and may even
disturb) the evaluation of the Metric. You should then apply
a filter such as


      RescaleIntensityImageFilter

or

       IntensityWindowingImageFilter

for preprocessing the images.

Note that these filters (and its parameters) bring uncertainty
to your comparision of Image metrics. For the sake of fairness
you probably want to apply *exactly* the same preprocessing to
the image that are fed into all your registration metrics.


Note that at the end, any comparision of Algorithms is pointless
and useless if you dont' provide the entire set of material that
you used for your comparision. That includes:

       - Source code
       - Input images
       - Full sets of parameters


Only in this way, other people will be able to repeat your
evaluations and tweak them in different ways. The fact that
each metric has many parameters makes very difficult (if not
impossible) to define a "fair" comparison. For example, you
are selecting for Viola Wells parameters such as :

     - Number of Bins
     - Number of Samples
     - Standard Deviations

Changes in any of those parameters will result in dramatic
changes on the outcome of the Metric, and therefore will chage
how this metric perform face to other metrics.


Conclusions of the sort:

               "Metric A is better than Metric B"

are useless and worst of all: misleading.


They are only of interest for writing papers in the Dark Side
of the current publishing system where reproducibility is not
supported or even welcomed, and where conclusions are not derived
from ones' own experience but from subjective judgement, such as
the ones provided by the decadent peer-review system.


Unfortunately, those practices still percolate the entire community
of medical image processing.



    Regards,




        Luis




-----------------------------
Jef Vandemeulebroucke wrote:

> Hi,
>  
> I am testing several mutual information metrics, plotting their behavior.
> Among the metrics are the two based on histograms:
>  
> MutualInformationHistogramImageToImageMetric
> NormalizedMutualInformationHistogramImageToImageMetric
>  
> Do these metrics give better results when the images have been 
> normalised, as it is for the Viola-Wells implementation of MI, or is 
> this of no importance?
>  
> Thank you,
>  
> Jef
> 
> 
> ------------------------------------------------------------------------
> 
> _______________________________________________
> Insight-users mailing list
> Insight-users at itk.org
> http://www.itk.org/mailman/listinfo/insight-users