[Insight-users] GPU Demons registration in ITK 4.2

[Gareth Price]

Hi there,

 

I was wondering if someone might be able to offer advice regarding the
new ITK 4.2 GPU implementation of the demons algorithm.

 

I have an application using the standard CPU implementation to register
3D CT images (512x512x~70 voxels). This behaves perfectly when
registering synthetic deformations to their original image. I typically
use a smoothing kernel standard deviation of 10.0 with no update field
smoothing. With my images, the metric reduces exponentially as expected
from ~8000 to 350 over 50 iterations with the RMS update field change
showing a similar but slower trend from ~0.08 to 0.04. After 50
iterations the solutions have made good progress towards the original
images. 

 

Porting this code to the GPU implementation following the examples in
the itkGPUDemonsRegistrationFilterTests (i.e. swapping the internal
image type for GPUImageType and GPUDemonsRegistrationFilter for
DemonsRegistrationFilter) provides an immediate ~x15 increase in speed
on my machine (GeForce GTX480). However using the same input data and
smoothing parameters, the metric and RMS update field trends, although
still exponential show changes from ~8000 to 1500 (metric) and ~0.3 to
0.2 (update field). The result is that the deformation field produced is
much too large. Changing smoothing parameters does not change this
behaviour, whilst enabling update field smoothing crashes the process
(in GPUDemonsRegistrationFilter::ApplyUpdate). It is very strange that
the similarity metric is continuing to reduce as the deformation field
pushes the moving image through and past its correct fixed image
alignment.

 

Would someone be able to tell me if this is expected behaviour/are known
bugs, or offer advice if I am using the GPU registration filter
incorrectly?

 

Many thanks, Gareth

 



****************************************************************
This e-mail and any files transmitted with it are confidential
and solely for the use of the intended recipient. If you have
received this e-mail in error you should not disseminate,
distribute or copy it. Please notify the sender immediately and
delete this e-mail from your system.
****************************************************************

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.itk.org/pipermail/insight-users/attachments/20120719/e9d8c544/attachment.htm>