Hi,
You are right that we reverse signs on similarity functions to turn them into cost functions.
We also drop some terms when they are constants (e.g. the entropy of the reference image, which is not transformed) since they do not affect the optimisation of the transformation parameters. And sometimes overall scaling constant are also left out. So all of these factors will lead to values that might not be in the range that you are expecting. Nonetheless, the function serves the purpose of determining the best transformation parameters just as well.
I hope that helps.
All the best,
Mark
> On 6 Apr 2018, at 15:05, Andrew Davis <[log in to unmask]> wrote:
>
> Hello image registration experts,
>
> I have a question about the mutual information cost metric produced by FLIRT. When I attempt to measure the mutual information of two images using the `measurecost1.sch` schedule file, I get a negative value. I had thought that mutual information was always a positive value. Can someone please point to a reference for the definition of mutual information used in FLIRT?
>
> For instance, running this command:
>
> flirt -in struct_brain -ref gre5-mag -cost mutualinfo -schedule $FSLDIR/etc/flirtsch/measurecost1.sch
>
> Produces this output:
>
>> -0.271626 0.000000 -0.000000 0.000000 0.000000 0.000000 0.000000 1.000000 1.000000 1.000000 0.000000 0.000000 0.000000
>>
>>
>> Final result:
>> 1.000000 0.000000 0.000000 0.000000
>> 0.000000 1.000000 0.000000 0.000000
>> 0.000000 0.000000 1.000000 0.000000
>> 0.000000 0.000000 0.000000 1.000000
>
> Which I take to mean that flirt is reporting a mutual information value of -0.271626. Also, using the normmi cost function produces a value of -1.08389, whereas I expected a value between 0 and 1.
>
> Yours,
> Andrew Davis
|