[Insight-users] Starting point for registration

[rieck at stud.ntnu.no]

I want to register two 3D volumes of MR scans of the head/brain, A and  
B. But I have a set of volumes
(Bs) i can register A with. I suspect the result will be better/compute
faster if I found
the already closest one in B to A and use this. Is there any similarity
measure/things to look for that would indicate a better registration?
I am not sure if I will use Deformable Registration or not.

At the moment i was thinking of using the bounding box of the head inside
the volume to rank them initially, and use a similarity measure [1] (if I
can get it to work in 3D) of the 2-3 best.

Another possibility (that handles volumes of different modes) is to use
one of the Mutual Information classes in ITK. (Possibly after the bounding
box ranking)

Can you think of other options? Is it even any point in doing this?

Also, do anybody know of a good method of evaluating the result of a
segmentation in 3D (given a gold standard). Is just counting false
positives/negatives sufficient? My initial search on scholar.google did
not give me to much to work with. :(



[1] Zamperoni, Starovoitov 1995. How dissimilar are two gray-scale images,
Proceedings of the 17 DAGM Symposium, Springer, Berlin pp 448-455.


best regards,
Christian Marshall Rieck