[Insight-users] What is best registration setting for binary image matching?

[Quy Pham Sy]

Thanks guys,

I check out MatchCardinalityImageToImageMetric (use exmaple in the
guide book), it's significantly faster than mean squares metric.
According to ItkSoftwareGuide, MatchCardinalityImageToImageMetric does
not require "analytical derivatives" of its cost function so they use
AmoebaOptimizer. it is  fast though, but AmoebaOptimizer just apply to
unimodal function. I try it and the result alway fall to a local
optimum value (which is incorrect). So question is

---> Which type of Optimizer should be use with MatchCardinality?

Sorry if bothering :)

quyps,

2009/10/16 Dan Mueller <dan.muel at gmail.com>:
> Hi Quyps,
>
> The following metrics are designed for binary/label images:
>    itk::MatchCardinalityImageToImageMetric
>    itk::KappaStatisticImageToImageMetric
>
> Computation of these metrics is relatively simple (ie. they should be
> faster than mean squares).
>
> Also, for performance make sure you configure/compile your application
> for optimized release mode.
>
> Hope this helps.
>
> Regards, Dan
>
> 2009/10/15 Kishore Mosaliganti <kishoreraom at gmail.com>:
>> If its a binary image, you can downsample it quite a bit. It won't
>> affect your registration transform.
>>
>> Kishore
>>
>> On Thu, Oct 15, 2009 at 11:13 AM, Quy Pham Sy <phamsyquybk at gmail.com> wrote:
>>> hi all,
>>>
>>> I need to match two binary images. first image actually is a subset of
>>> second one.
>>> The corresponding transform  between two images is composition of
>>> translation, isotropic scaling, and centered rotation.
>>> I use CenteredSimilarity2DTransform for my transformation type.
>>>
>>> In my current setting, i use: RegularStepGradientDescentOptimizer,
>>> MeanSquaresImageToImageMetric,
>>> NearestNeighborInterpolateImageFunction.
>>>
>>> But it takes more 6 minutes to finish a registration process.i think
>>> it is long for two binary images.
>>>
>>> Anyone know what type of setting (metric, optimizor...etc) i should
>>> use in this case?
>>>
>>> Thanks.
>>> Quyps
>



-- 
Pham Sy Quy
HCI Lab, Advanced Fusion Technology Department,
Room 1211, New Millennium Building
Konkuk University, Seoul, Korea
Mobile: +82-10-9800-8104