[Insight-users] about MIThinPlateSplineRegistrator

[Luis Ibanez]

Hi Zhang,

The choice of using the Source landmarks as
parameters is totaly arbitrary.

We can equally go with using the Target Landmarks,
or even using both.

Note however that this is independent of how the
mututal information metric sample the images
in order to estimate the joint probabilities.
It is in fact independent of any metric used
for registration.

The MI sampling points are not related at all with
the Source and Target landamarks, except for the
fact that the landmarks define the transform,
and the pairs of sampling points must be one
the map of the other under the transform.

A bit of experimentation with the ThinPlateSpline
example could help to illustrate the implications
of choosing the paramters as:

1) The coordinates of the Source landmarks
2) The coordinates of the Target landmarks
3) The coordinates of both Source and Target landmaks

Since the Spline transform is in fact computed
using the vectors relating one Source landmark
with one Tartget landmark, the real thing to
modify is this relative position and that can
be done equally well by changing the source
landmarks or the target landmark. (e.g. moving
one source landmark to the right is equal to
moving all the other target landmarks to the
left.

Option (3) is interesting if we consider this
previous case, since the optimizer will probably
succed in moving the only one source landmak
instead of the N-1 target landmarks.  However
the price is to double the dimension of the
parameter space, and henceforth double the
time for registration.


It could probably be convenient to add an option
controlled by an enum, so the user can externally
select how the parameters should be managed.
The options would be the three enumerated above.

That should be relatively simple to implement.

Please let us know if you find this
alternative reasonable.

Thanks

Luis


===================================================

zhangzhijun wrote:

> hi Luis:
> 
>          I am wondering whether choosing of source landmark as parameters
> will be ok,
> the mutual information implemented in ITK using the stochastic gradient
> descent to find
> the optimal. The entropy and the joint entropy are estimated by using 2
> groups of sampling
> points and for each iteration the sampling points in source and target image
> are chozen arbitrarily
> and are not same, so How could the target landmark be fixed and the source
> landmark
> be adjusted?
> 
> regards,
> zhang zhijun
> 
>