[Insight-users] some questions about Demons filter

[Torsten Rohlfing]

Hi Borja --

I did some experiments recently where I compared the performance of the 
Demons filter to that of the Curvature PDE filter and my own private 
implementation (not in ITK) of Rueckert's B-spline algorithm. The 
experiment used data from the Internat Brain Segmentation Repository 
(IBSR) to quantify overlap of anatomical structures in MR brain images 
after nonrigid inter-subject registration. Basically, say you have MR 
image I1 and corresponding segmentation S1, and likewise image I2 and 
segmentation S2. Then use a registration algorithm to compute the 
transformation from I1 to I2. Then you can pull back the segmentation S2 
onto I1, where you can compare the pulled-back segmentation with the 
(supposedly) correct one S1.

When I did this and computed the recognition rates (i.e., the number of 
correctly labeled pixels), this is what I got:

B-spline: 0.96759
Demons: 0.96689
Curvature: 0.96471

You see that the differences seem marginal, but for what it's worth, 
Demons seem to work slightly better than the Curvature registration 
algorithm, and slightly worse than my version of "The Rueckert."

As for the parameters: I ran Demons and Curvature in a four-level 
multi-resolution scheme using ITK's multi-resolution registration 
framework. The Rueckert algorithm as run with an initial 40 mm control 
point grid that was refined 4 times by a factor of 2, down to 2.5 mm 
final spacing.

TR
-------------- next part --------------
A non-text attachment was scrubbed...
Name: torsten.vcf
Type: text/x-vcard
Size: 366 bytes
Desc: not available
Url : http://public.kitware.com/pipermail/insight-users/attachments/20051102/40f7ca38/torsten.vcf