Doubts about evaluating metric


I have a very confused question. When calculating the similarity of two identical images, the value obtained should be 1 or greater than 1 (because I don’t know whether the value is normalized or not), but a very strange problem arises. The value obtained is more than - 0.8. There are two problems: (1) the problem of negative value. I have consulted the documents and papers, and understand that it is caused by the definition. (2) If the value is more than 0.8, I don’t know how to understand it. The metric I use is “jointhistogrammutualinformation”.

And when I rotate the registered floating image 1 ° in one direction (or 1 mm translation), I find that the calculated value is better than the registered value. Is this the problem of registration algorithm?

The registration framework is set up to always minimize the metric (think mean square error), so mutual information metrics report negative values (the more negative, the better).

Thank you for your answer.
I understand that mutual information value is negative in simpleitk. The smaller the value, the better.

But what I don’t understand is: (1) the parameters after registration are optimized, which should be the best in theory, right? But when I add 1 ° or 1 mm to this parameter, the result is better, which puzzles me. (2) When I use “join this diagram information”, I use two identical graphs to measure the similarity without registration. I don’t know if the value is normalized, so the value should be no more than - 1, but the result is not, the value is - 0.8.

Value -0.8 is worse than -1.0, because less is better, and -1.0<-0.8. I guess that MI’s range is [-1.0, 0.0].

Yes, that’s right.But the graphs I used are the same, so I think the value is - 1, not - 0.8 .(use your range [- 1,0])