[Insight-users] MattesMutualInformation metric: Fixed and moving image desired features

[Oscar Esteban Sanz-Dranguet]

Hi,

I'm trying to perform intra-subject rigid registration between MR & 
SPECT images. I would like the MR to be the fixed image, BUT I've found 
that if I use the SPECT as fixed, the registration is more quick and 
accurate. Why does it happen?

The images are as follows:

- SPECT: 128x128x43,
               3,32237mmx3,32237mmx3,32237
               less than 255 levels of gray

- MR: 224x240x256
          0,85mmx0,9mmx0,85mm
          512 levels of gray


Image preprocessing:
- I use a IntensityWindowingImageFilter to rescale the intensities to 
the range 0,255.
- I use a DiscreteGaussianImageFilter with deviation 2.0 to the two 
images, but I've tried using a lot of different values and combinations. 
It appears to be the same
- I use ScalarImageToHistogramGenerator to find the first maximum of the 
two images histogram. With it, I use the metric's  member 
SetFixedImageSamplesIntensityThreshold for the fixed image and I remove 
the background of the moving image with ThresholdImagefilter. The values 
used for the two thresholds is the value of the first maximum + 10% of 
overall frequencies.
- Resampling SPECT to MR's spacing and size improves the registration, 
when using MR as fixed and adequate values of optimizer scales (they are 
in a very tiny range)

Initialization:
- Using Moments, the SPECT image starts a little lower (10mm) on the Z axis

Metric:
- 70 bins, 10000 samples

Optimizer:
- SPSA or Gradient Descent. I use a visualization pipeline with VTK to 
test that the scales and the step size make sense.


Thanks in advance. Cheers.

-- 
Oscar Esteban
BIT - UPM (http://www.die.upm.es/im/)
+34 913 366 827 ext.4248