[Insight-users] Re: Registration question - optimizer

[Luis Ibanez]

Hi Mary,

You should start by selecting an appropriate
initialization for the TranslationTransform.

Please look at the ITK Software Guide for
instructions on how to initialize the Transform.

A typical approach is to compute a translation
that will overlap the center of both images.
This make sense only if the anatomical structures
that you want to register are also well centered
on each of the two images.

It is now worth to spend time yet on fine tunning
the optimizer if you dont' have a reasonable
initialization yet.

In order to test how good the initialization is
you can simply resample the Moving image in the
coordinate frame of the Fixed image using the
initial transform. Then overlap the two images
(with the checkerboard filter, for example) and
visually evaluate how well they fit.



    Regards,


       Luis



----------------
Mary Chou wrote:
> We are trying to register images using the TranslationTransform, 
> GradientDescentStepOptimizer, and the Mean Square Image to Image metric, 
> but are having a problem with the convergence of the method, in that the 
> metric values do
> not seem to converge to a minimum at all.  The original images are 0 - 
> ~6500 in intensity value range.
> 
> We have tried registration using the ITK Examples file 
> ImageRegistration1.cxx and have tried adjusting the values for the 
> Optimizer, maximum step length and minimum step length.  Any suggestions 
> for optimizer parameter selection?
> 
> Please find the data (original images) attached.
> 
> 
> //  optimizer->SetMaximumStepLength( 4.00 ); 
> //  optimizer->SetMinimumStepLength( 0.01 );
>   optimizer->SetMaximumStepLength( 0.01 ); 
>   optimizer->SetMinimumStepLength( 0.0001 );
> 
> 0 = 2.72438e+006 : [-2.08739, -3.41216]
> 1 = 1.73376e+006 : [-1.98374, -1.41484]
> 2 = 2.01186e+006 : [-2.5951, -2.2062]
> 3 = 1.70106e+006 : [-3.36987, -2.83844]
> 4 = 1.6264e+006 : [-3.03848, -2.46403]
> 5 = 1.64197e+006 : [-2.67566, -2.11999]
> 6 = 1.70971e+006 : [-2.86588, -2.28223]
> 7 = 1.66576e+006 : [-3.04977, -2.45159]
> 8 = 1.64041e+006 : [-2.95655, -2.36831]
> 9 = 1.65473e+006 : [-2.99801, -2.41508]
> 10 = 1.65153e+006 : [-3.01273, -2.47582]
> 11 = 1.64687e+006 : [-2.99193, -2.4525]
> 12 = 1.6475e+006 : [-2.96613, -2.47014]
> 13 = 1.64447e+006 : [-2.96402, -2.50132]
> 14 = 1.64206e+006 : [-2.98953, -2.51937]
> 15 = 1.64255e+006 : [-3.01521, -2.53718]
> 16 = 1.64318e+006 : [-3.04244, -2.55252]
> 17 = 1.6372e+006 : [-3.07013, -2.567]
> 18 = 1.63179e+006 : [-3.09843, -2.58026]
> 19 = 1.62691e+006 : [-3.12752, -2.59168]
> 20 = 1.62251e+006 : [-3.15762, -2.60005]
> 21 = 1.61852e+006 : [-3.18874, -2.6029]
> 22 = 1.61484e+006 : [-3.2191, -2.5955]
> 23 = 1.61146e+006 : [-3.24348, -2.57595]
> 24 = 1.60868e+006 : [-3.26375, -2.55217]
> 25 = 1.60643e+006 : [-3.28371, -2.52812]
> 26 = 1.60444e+006 : [-3.30364, -2.50405]
> 27 = 1.60269e+006 : [-3.32351, -2.47993]
> 28 = 1.60117e+006 : [-3.31086, -2.47077]
> 29 = 1.60208e+006 : [-3.29818, -2.46164]
> 30 = 1.60315e+006 : [-3.28548, -2.45253]
> 31 = 1.60439e+006 : [-3.27277, -2.44345]
> 32 = 1.60579e+006 : [-3.26002, -2.43441]
> 33 = 1.60736e+006 : [-3.24725, -2.42541]
> 34 = 1.6091e+006 : [-3.23444, -2.41646]
> 35 = 1.61101e+006 : [-3.22159, -2.40757]
> 36 = 1.61309e+006 : [-3.20869, -2.39876]
> 37 = 1.61534e+006 : [-3.19573, -2.39003]
> 38 = 1.61777e+006 : [-3.18269, -2.38143]
> 39 = 1.62038e+006 : [-3.16955, -2.37297]
> 40 = 1.62317e+006 : [-3.15627, -2.36474]
> 41 = 1.62614e+006 : [-3.1428, -2.35682]
> 42 = 1.6293e+006 : [-3.12903, -2.34942]
> 43 = 1.63265e+006 : [-3.11478, -2.34303]
> 44 = 1.63618e+006 : [-3.09962, -2.33923]
> 45 = 1.63979e+006 : [-3.08574, -2.34639]
> 46 = 1.64205e+006 : [-3.08088, -2.36125]
> 47 = 1.64154e+006 : [-3.0654, -2.36335]
> Result =
>  Translation X = -3.0654
>  Translation Y = -2.36335
>  Iterations    = 49
>  Metric value  = 1.64479e+006
>