[Insight-users] Can a metric be negative in ITK?

[Christopher Kappe]

Thanks for the hint with the multi resolution registration - I'll consider 
this. 
And just to clearify this: I did some "deeper" research and found it 
explicitely stated somewhere that Mattes Mutial Information Metric has a 
negative outcome in ITK. And it is 0 (zero) for two totally different images. 
So as already assumed one wants a high (negative) number for matching images. 


Am Montag 05 März 2012, 12:23:12 schrieb Rick Frank:
> I believe that is the case. You are trying to minimize a metric, which
> is going to be where it's derivative at its global minimum is 0,
> Not necessarily the value of the metric.
> 
> Also, one must be aware that typically MI has a very "jaggy" metric
> trace, which will create local minima. One needs an optimizer that can
> Jump out of the local minima, or use a pyramid of resolutions,  and/or,
> you can jitter pixels (as per Joachim Hornegger's "Korea"
> talk/presentation) to try to smooth the metric. I have not tried the
> latter but he claims it will smooth the resulting trace making
> optimization less likely to stick in local minima.
> 
> 
> You can find his lectures here
> 
> http://www5.informatik.uni-erlangen.de/en/our-team/hornegger-joachim/lec
> tures/
> 
> Rick
> -----Original Message-----
> 
> Date: Sun, 04 Mar 2012 16:56:54 +0100
> From: Christopher Kappe <cpkappe at gmx.de>
> Subject: Re: [Insight-users] Can a metric be negative in ITK?
> To: insight-users at itk.org
> Message-ID: <1377188.YjJcHt4Ey7 at chrisbook.site>
> Content-Type: text/plain; charset="iso-8859-1"
> 
> So it's absolutely normal that this metric returns negative values and
> it's 
> okay when they get even "more negative" during the registration process.
> (?)
> 
> Paul Muetze
> 
> Am Freitag 02 M?rz 2012, 12:04:01 schrieb Rick Frank:
> > "High" negative  is what you want from this metric, as I understand
> it.
> > 
> > Rick
> >
> > 
> > Hello everybody, 
> > 
> > I'm new to this list. 
> > I'm trying to do registration with two 3D images. I basically used the
> > ImageRegistration8.cxx example (using itk 4.0 btw) but with
> > MattesMutualInformationMetric instead of MeanSquaresMetric. 
> > And, well, I got output like this from the observer: 
> > 
> > 0   -0.00130253  
> > 1   -0.000181399 
> > 2   -0.00019049   
> > 3   -5.82618e-06
> > 4   -6.76423e-06 
> > 5   -1.81517e-05  
> > 6   -1.78196e-05   
> > 7   -7.79246e-06  
> > 8   -4.0121e-06
> > 9   -3.87818e-06 
> > 10   -1.20782e-05   
> > 11   -2.655e-06   
> > 12   -1.13154e-05  
> > 13   -7.53967e-06  
> > 14   -8.0584e-06  
> > 15   -3.52788e-06 
> > 16   -2.35729e-06
> > 17   -5.10754e-06
> > 18   -6.61478e-05
> > 19   -0.000101335
> > ...
> > 49   -0.000806308 
> > Optimizer stop condition: VersorRigid3DTransformOptimizer: Maximum
> > number of iterations (50) exceeded.
> >           Probe Tag    Starts    Stops             Time (s)
> >         Registration           1            1           4509.42
> > 
> > The thing is, it's not even converging (at least not towards 0).
> > And is it okay to get a negative metric as result in the first place?
> > 
> > Best
> > Paul Muetze