[Insight-users] A Problem for MI Registration

[Luis Ibanez]

Hi Weihua,

The two binary images that you are using are particularly
ill-suited for Mutual Information.

You may want to use Normalized Mutual Information for this
particular pair of images, or use a more reasonable metric,
such as MeanSquares and blurr the input images as Julien
suggested.

I would assume that you are using these two binary images
just as a toy example before you go for registration on
real medical images. If this is the case, I would strongly
suggest you to use synthetic images that are more similar
to medical images, e.g. blurred blobs with some texture.


    Regards


      Luis



-----------------------
Julien Jomier wrote:

> Hi Weihua,
> 
> Registration often requires tuning the parameters depending on the 
> optimizer, metric, and the nature of the images.
> 
> You need to play a little bit with the parameters of the registration.
> Here a list of what you can do to adjust the ImageRegistration2 example.
> 
> 1) Optimizer
>   a) Increase the number of iterations so the registration doesn't stop 
>        too quickly.
>   b) Increase/Decrease the learning rate. The value of the gradient is 
> multiplied by the learning rate.
> 
> 2) Metric (MI)
>   a) Adjust the ImageStandardDeviation for both the moving and the fixed 
> images.
>   b) Increase the number of Spatial Samples to achieve better convergence.
> 
> 3) Pre-Processing
>   In this example the preprocessing step blurs the input images. The 
> problem you are having whith your binary images and the gradient 
> optimizer is that if the initial translation is too far from the answer, 
> the optimizer cannot follow the gradient, because the gradient is flat.
> 
> That said, using a learning rate of 20 and a number of Iterations of 
> 2000 and blurring the images by a factor of 3 (instead of 2) leads to a 
> convergence of  Translation X = -10.4966, Translation Y = -2.33114. I'm 
> confident you can adjust the translation in Y by increasing the number 
> of spatial samples and/or adjust other parameters.
> 
> Hope this helps,
> 
> Julien
> 
> Weihua Xiong wrote:
> 
>> Hi:
>>
>>
>> I made two images with only a WHITE PATCH in them. The White patch 
>> between
>> them are 10 pixels along x direction. When I used the 
>> 'ImageRegistration2'
>> to find the transfomation with MI, the result is
>>
>>
>> Result = 
>>  Translation X = -2.42836
>>  Translation Y = -0.352346
>>  Iterations    = 200
>>  Metric value  = 0.0369613
>>
>>
>> Do you what is the reason? The fixed image and the moving image are
>> attached.
>>
>>
>>
>> Thanks
>>
>>
>>
>> ---------------------------------------------
>>
>> Weihua Xiong PH.D.
>>
>> Computational Color Vision Lab
>>
>> Simon Fraser University
>>
>> 604-630-6060
>>
>> www.cs.sfu.ca/~weihuax/personal
>>
>> --------------------------------------------
>>
>>
>> ------------------------------------------------------------------------
>>
>>
>> ------------------------------------------------------------------------
>>
>>
>> ------------------------------------------------------------------------
>>
>> _______________________________________________
>> Insight-users mailing list
>> Insight-users at itk.org
>> http://www.itk.org/mailman/listinfo/insight-users
> 
> _______________________________________________
> Insight-users mailing list
> Insight-users at itk.org
> http://www.itk.org/mailman/listinfo/insight-users
> 
>