[Insight-users] Automatic Landmark localization

[Lars Nygard]

Hi all,

I'm working on an automatic landmark localisation algorithm in MR images. Im trying to
locate the nasion (saddle point between the eyes) and two points on the ears.
My algorithm is based on registration to an atlas. First I manually locate the landmark
and then I register to the atlas (icbm_152). I use two images to define the landmark in the atlas and
use the third image to measure the accuracy of the algorithm by applying the inverse transformation
 to the landmark in the atlas.
I've tried a two step registration (rigid + affine) and the euclidean distance between manual and automatic
localization is about 10mm. 
Currently I'm trying the BSplineDeformableTransform but the results get worse to about 20mm. I use the
InverseDeformationFieldImageFilter and everything seems to be working fine but I don't understand why
the result get worse. I've tried several setting but am at loss at the moment.
Anybody any tips for fine-tuning a deformable transform? The algorithm takes a long time, especially inverting
the deformation field (for data info see below). How can I speed up the algorithm? Any help or ideas will be appreciated.
regards,
Lars

fixed
ObjectType = Image
NDims = 3
BinaryData = True
BinaryDataByteOrderMSB = True
ElementSpacing = 1 1 1
DimSize = 181 217 181
Position = -90 -126 -72
ElementNumberOfChannels = 1
ElementType = MET_UCHAR
ElementDataFile = icbm_avg_152_t1_tal_lin.raw

moving
ObjectType = Image
NDims = 3
BinaryData = True
BinaryDataByteOrderMSB = True
ElementSpacing = 1 1 1
DimSize = 256 256 180
Position = -127 -112.92 -82.4405
ElementNumberOfChannels = 1
ElementType = MET_UCHAR
ElementDataFile = n474_mra.raw


__________________________________________________
Bruker du Yahoo!?
Lei av spam?  Yahoo! Mail har den beste spambeskyttelsen 
http://no.mail.yahoo.com 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.itk.org/pipermail/insight-users/attachments/20080612/31657b59/attachment.htm>