[Insight-users] Registering US -> CT
Luis Ibanez
luis.ibanez at kitware.com
Thu Sep 23 18:22:54 EDT 2004
Hi Neilson,
Actually, there is another option.
You could calculate 2D derivatives in the Ultrasound Images,
take the 2D vectors and map them into their real position and
orientation in 3D space by taking into account the spatial
position and orientation of the US slice.
You still need to define a width on the US slice in order to
determine which points from the 3D CT are considered to be
interacting with the US slice.
In this way you simply use a GradientRecursiveGaussianImageFilter
instantiated in 2D for computing the gradients from the US.
Regards,
Luis
----------------------------
N.E. Mackenzie Mackay wrote:
> Hey Luis,
>
> If I did do this would this be the only way to actually calculate
> the derivative? I was looking through the
> itkNormalizedCorrelationPointSetToImageMetric and it looks like the
> points that are on the image are the only ones taken into account when
> calculating the gradient.
>
> In my original code with Mutual information I had the Ultrasound
> image as the fixed image so that the gradient would be calculated using
> the 3D CT. For any PointSetToImageMetric I can't set the US image as
> the fixed image. Making the gradient only calculated using the 2D US.
> Even if I applied the blurring would this make it possible for the
> Gradient to be calculated properly?
>
> Thanks,
> Neilson
>
> On Sep 19, 2004, at 7:51 PM, Luis Ibanez wrote:
>
>>
>> Hi Neilson,
>>
>> No,
>> if you read a 2D image and you apply a 3D blurring, what you
>> will get is an Exception thrown. A 2D image is considered
>> to be a degenerate 3D image and therefore it is not suitable
>> for being the input of a 3D filter.
>>
>> What you can do, however, is to read the 2D image, declare
>> a 3D image with thickness 5 pixels (or 2N+1...) copy the
>> slice that you read into the middle slice of the 2N+1
>> image. Then, apply a blurring filter.
>>
>> You can set the pixel spacing Dz in this new 3D image in such
>> a way that Dz * (2N+1) = the physical thicknes of the sensitive
>> plane of your Ultra Sound acquisition probe.
>>
>> In that way this new 3D image will be a better representation
>> of the physical reality that your Ultra Sound 2D image captured.
>>
>>
>> Regards,
>>
>>
>> Luis
>>
>>
>> ---------------------------------
>> N.E. Mackenzie Mackay wrote:
>>
>>> That sounds like a good idea.
>>> If I read in the 2d image as 3D and applied a blurring filter would
>>> that cause a 3D blurring ( blur the pixels outside of the plane )?
>>> If so, I could set the "thickness" by the setting the radius of the
>>> blurring mask.
>>> Just a thought.
>>> Neilson
>>> On Sep 16, 2004, at 6:47 PM, Luis Ibanez wrote:
>>>
>>>>
>>>> Hi Neilson,
>>>>
>>>> That's a good point. You are right,
>>>> in this case, since you have scattered slices
>>>> from Ultra Sound it is not possible to make sure
>>>> that every point will be inside an image.
>>>>
>>>> One option could be to associate a "thickness"
>>>> to every US slice, you can probably figure out
>>>> one that makes sense from the point of view of
>>>> the physical acquisicion process.
>>>>
>>>> That thickness could be used for defining
>>>> regions of space where a point will be considered
>>>> to be "inside" one of the US images.
>>>>
>>>> The more US image you have, the more chances
>>>> there are that this could lead to a reasonable
>>>> registration.
>>>>
>>>> Note that this requires you to do more modifications
>>>> on the ImageMetric class.
>>>>
>>>>
>>>>
>>>>
>>>> Regards,
>>>>
>>>>
>>>> Luis
>>>>
>>>>
>>>>
>>>> --------------------------------
>>>> N.E. Mackenzie Mackay wrote:
>>>>
>>>>> I was thinking the same thing.
>>>>> The only thing I was worried about is using the method in 3D. If
>>>>> some of the points don't map onto the the US image will the
>>>>> registration method ignore those points or will it throw an error?
>>>>> On Sep 14, 2004, at 10:33 PM, Luis Ibanez wrote:
>>>>>
>>>>>>
>>>>>> Hi Neilson,
>>>>>>
>>>>>> MutualInformation is mostly a region-based image metric.
>>>>>> This means that its value gets better when the overlap
>>>>>> between matching regions of the image is large. Mutual
>>>>>> Information is not particularly well suited for matching
>>>>>> thin structures since thir random sampling is unlikely
>>>>>> to select many pixels belonging to those structures.
>>>>>>
>>>>>> In that sense you probably shouldn't expect much from Mutual
>>>>>> Information for registering Bone, since bone structures are
>>>>>> mostly shell-like and they don't fill large regions of space.
>>>>>> E.g. large bones have their layers of cortical bone with high
>>>>>> calcifications but their width usually cover just a couple of
>>>>>> pixels in a CT scan.
>>>>>>
>>>>>> Given that you seem to have segmented the bone from the CT scan,
>>>>>> it is probably worth to try a Model-to-Image registration approach.
>>>>>> This can be done by taking points on the surface of your bone
>>>>>> segmentation, and/or from a band around that surface, and using
>>>>>> them to match the intensities (and structure) of the same bone as
>>>>>> seen in the UltraSound images.
>>>>>>
>>>>>> Could you post a couple of the US images ?
>>>>>>
>>>>>> (e.g. you could put them in www.mypacs.net and let us know their
>>>>>> image ID).
>>>>>>
>>>>>>
>>>>>> Depending on how the bone structures look like on the US image
>>>>>> there may be different possible metrics to try in a PointSet to
>>>>>> Image registration.
>>>>>>
>>>>>>
>>>>>> BTW, when you start working in 3D, don't attempt to use Rigid
>>>>>> transforms until you have manage to tune all the other parameters
>>>>>> of the registration to work with simple translation transforms.
>>>>>> It is more effective to deal with a single issue at a time.
>>>>>>
>>>>>>
>>>>>>
>>>>>> Regards,
>>>>>>
>>>>>>
>>>>>> Luis
>>>>>>
>>>>>>
>>>>>>
>>>>>> -------------------------------
>>>>>> N.E. Mackenzie Mackay wrote:
>>>>>>
>>>>>>> Hi,
>>>>>>> I have tried for the last while to get a single ultrasound
>>>>>>> image to register to a CT volume. Specifically try to get the
>>>>>>> bone of an ultrasound image and bone of the CT to register
>>>>>>> together. Up to now I am having quite some trouble.
>>>>>>> I have been able to segment the bone from CT and give an estimate
>>>>>>> on where the bone is in the ultrasound. I am now trying to
>>>>>>> register those two images.
>>>>>>> This is what I am using:
>>>>>>> MattesMutualInformationImageToImageMetric - decided to use this
>>>>>>> because the registration was of two different modalities.
>>>>>>> Couldn't use feature registration becuase to hard to segment
>>>>>>> ultrasound correctly
>>>>>>> - using 50 bins and %20-%100 of samples still doesn't give
>>>>>>> adequate results.
>>>>>>> linearInterpolateImageFunction
>>>>>>> RegularSetGradientDecentOptimizer
>>>>>>> Euler3DTransform
>>>>>>> Both images ( 3D CT and US ) are normalized before the registration.
>>>>>>> I have used a maximum step ranging from 0.5-6. And a min step of
>>>>>>> 0.005 - 1.
>>>>>>> I have a good initial guess ( maximum 2cm away from correct with
>>>>>>> about 0- 30degrees of rotation)> I tested out the registration
>>>>>>> method in 2D and have had success. When I use the exact same
>>>>>>> variables applied in 3D the registration is poor.
>>>>>>> Does anyone have any suggestions? I would be happy to provide a
>>>>>>> couple of images or actual code to show you what I am dealing with.
>>>>>>> Neilson
>>>>>>> _______________________________________________
>>>>>>> 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
>>
>>
>>
>>
>>
>>
>
> _______________________________________________
> Insight-users mailing list
> Insight-users at itk.org
> http://www.itk.org/mailman/listinfo/insight-users
>
More information about the Insight-users
mailing list