If you are using Opencv  You have inbuilt fuctions to

or you can calculate 2d correlation to get the best match

4. Or You my put Your segmented images into clusters. Find the centers of each clusters. And then compare the new images with this centers.

Scale Invariant Feature Transform (SIFT) is the best nowadays, SURF is a semilar approach but faster. To match these features, there are different well-known similarity measurements.

is SURF or SIFT applied to medical images like CT scans or MRI

Try a new and promising (yet sufficiently simple) technique proposed in the paper

     Alexander V. Prolubnikov,

     An interval approach to pattern recognition of numerical matrices //

     Reliable Computing 19 (2013), Issue 1, pp. 107-119.

Direct link to the paper is

http://interval.louisiana.edu/reliable-computing-journal/volume-19/reliable-computing-19-pp-107-119.pdf

I think you can try ranking or re-ranking algorithms. They are specific descriptors and similarity functions, according to the kind of image you work.

Mr Marco Antonio Garcia de Carvalho can you  suggest to me some ranking algorithms for real images?

Thank you 

Yes. I recommend this paper:

Daniel C. Pedronette, Ricardo S. Torres. Image re-ranking and rank aggregation based on similarity of ranked lists. Pattern Recognition 46 (2013), 2350-2360.

Best regards!