[Insight-users] Model based registration for vessel axis extraction

[fjlb at duke.edu]

Hi,


I'm trying to modify the code in the model based registration example in the
user guide for my application. Basically I want to use this registration
framework to extract the axis of vessels in 2D and 3D. First I'm trying to do
it in 2D for simplicity. I have a fixed image (manually created phantom) that
is just a white curve on a black background that I blurred with a gaussian such
that the highest intensity pixels are in the middle of the curve. Since I want
to extract the axis of this curve I want my model to be represented by the Line
Spatial Object. I can create an initial estimate (line) that is close to the
curve and ideally the registration procedure will take my spatial object to the
center of the curve (or vessel).

The second part of the problem is that this curve moves and changes shape over
time (coronary artery's motion over cardiac cycle). The idea is to use the
results of the registration in the first frame as the initial estimate (line
spatial object) in the second frame (whose fixed image is the new curve ) and
so on to eventually have an optimized axis for each frame. By doing it this
way, we can track the motion of particular points in the line through time.

First of all, I want to know which combination of transform, metric, and
optimizer are the more suitable for my application? I have written a metric for
this application based on the external energy of 'snakes' (add the pixel values
for each point location - the highest value of this will be when all points of
the line are in the center of the curve).

I attempted to solve the problem with no luck using the line spatial object
along with the snakes metric, a Bspline transform, and the LBFGS optimizer. As
the registration procedure moves on, some control points in the Bspline grid
just blow up.


Any suggestions are more than welcome.


Thanks,


Federico