I have a problem concerning the measurement of similarity between two images (or matrices).
I was redirected here from the sci.stat.consult list (see https://groups.google.com/forum/?fromgroups=#!topic/sci.stat.consult/e6KzgKPY2V8).
I need a method to quantify similarity between two (same dimensions, grayscale) images. However, it is important that this method takes spatial relation/neighborhood into account.
An over-simplified example that however demonstrates why I am not really satisfied with sth like RMS (or any pixel-by-pixel comparison) is given here:
The method should be able to "show" that these images are quite similar:
While these are not:
Here is an example of two images (cardiac polar plots of the left ventricular myocardium) that I am in fact interested in:
From a doctor's perception consistency and agreement in the example above would be rather high.
I need a method to quantify consistency and agreement between two imaging modalities (MRI and PET) based on these polar plots. However, it is important that this method is quite robust against slight miss-alignements between those plots which definitely occur as the heart is steadily moving while we are doing image acquisitions.
Disclaimer: I am a radiologist (with some interest in statistics and quite some experience in bioinformatics). However, I am not a statistician. So please forgive any inaccuracies in my question. I'll do my best to put is as precise as possible.
Thank you very much in advance,