[Elastix] Groupwise nD+t B-splines for 4DCT of lungs
Einstein, Daniel R
Daniel.Einstein at pnnl.gov
Sat Feb 14 17:28:33 CET 2015
Thanks Geoff and Floris,
Good ideas both. I will give them a try and let you know what I find out.
From: Geoff Hugo <gdhugo at vcu.edu<mailto:gdhugo at vcu.edu>>
Date: Saturday, February 14, 2015 4:37 AM
To: "floris at isi.uu.nl<mailto:floris at isi.uu.nl>" <floris at isi.uu.nl<mailto:floris at isi.uu.nl>>
Cc: "Einstein, Daniel R" <daniel.einstein at pnnl.gov<mailto:daniel.einstein at pnnl.gov>>, "elastix at bigr.nl<mailto:elastix at bigr.nl>" <elastix at bigr.nl<mailto:elastix at bigr.nl>>
Subject: Re: [Elastix] Groupwise nD+t B-splines for 4DCT of lungs
Dan and Floris,
We have had good results with something similar to the suggestion in 2c on a similar problem. Set up your 4D image as a 4D volume as Floris suggested. Then make a 4D volume out of your TLC image by stacking this 3D image 't' times, and use this as the fixed image. You then perform registration using an existing non-groupwise metric, such as meansquares for CT. This results in a 4D transform mapping each time frame to the TLC image, which is what it sounds like you are after?
You can add temporal smoothing by controlling the bspline grid size in the time direction. If you add a regularization term such as the bending energy penalty, it may help to also enforce smoothness, although we have gotten good results without it. The other advantage to this approach is that many of the basic metrics are multithreaded, whereas the varianceoverlastdimension metric used in the group wise approach is not, and can be substantially slower.
Geoffrey D. Hugo, Ph.D.
Associate Professor, Radiation Oncology
Director, Medical Physics Graduate Program
Virginia Commonwealth University
(804) 628 3457
On Feb 13, 2015, at 8:01 AM, Floris Berendsen <floris at isi.uu.nl<mailto:floris at isi.uu.nl>> wrote:
That is an interesting application. I was not involved in the works you mention, however I can give my view based on experiences with these methods and the elastix framework.
Unfortunately, a set of 3d images must be provided as a 4D volume with its last dimension the image number or time index. In the current framework of elastix the groupwise registration is sort of hacked in, therefore it has these restrictions.
I think my answer is threefold.
2a): I would assume that more (independent) image data can provide a better registration (estimate) in general. A higher SNR is a bonus as well.
2b): In the current groupwise registration framework the resulting transformation is defined to point from a floating 'spatial average' image domain to each time frame domain. Since you would like to model all deformations in the space of the TLC image you would need to chain the resulting transformations to each time frame with the inverse transformation of the TLC. This can be done with some scripting.
As a downside of including the TLC image in the time stack you probably should not use any time domain smoothness for the registration, since that doesn't make sense for such a hybrid stack.
2c): Ultimately, you might want to adapt the groupwise framework such that you have a the TLC image as fixed image and the time stack as moving image. In that way you have the TLC image as a reference domain for modelling and the image similarity metric is still calculated in a groupwise fashion. You would need to write an adapted metric and possibly transformation for that, but the current framework allows this, I think.
Not sure about the details in the paper. I would say that the lung mask needs to be 4D (and possibly different for each 3d sub volume), such that the sampler knows what the valid samples are.
External tools/labour are needed to generate these label images. Delmon et al. use "Automated Segmentation of a Motion Mask to Preserve Sliding Motion in Deformable Registration of Thoracic CT" for that. I worked on an approach similar to that of Delmon et al. My souce code is not yet part of elastix, but the method is published so far is as:
Registration of organs with sliding interfaces and changing topologies
Combining sliding organs with a groupwise approach has come to my mind as well. I think it can be done, but requires bigger adaptations of the code and might be limited to certain setups. Both our slidng organ methods require a label image that is defined in the fixed domain. In principle this conflicts with the floating spatial domain, but 2c) would solve this.
On Thu, Feb 12, 2015 at 9:42 PM, Einstein, Daniel R <Daniel.Einstein at pnnl.gov<mailto:Daniel.Einstein at pnnl.gov>> wrote:
I am interested in applying a number of ElastiX 4D lung registration approaches to 11 dynamic rat lung images that are described in:
The parameter file (Par0012) available at http://elastix.bigr.nl/wiki/index.php/Par0012 seems to be what I am after.
Question 1): do the images input to the elastic command line need to be truly 4D, or can they be multiple 3D images?
elastix -f <dynamic nD+t image> -m <dynamic nD+t image> -p <par filename> -out <output dir>
Question 2): We have a higher-resolution, higher SNR static TLC image from the same animal upon which we would like to base a model geometry. Assuming both sets of images (4D + TLC) are up/down sampled to be equal, do you think there is any benefit to including the TLC image in the groupwise registration?
Question 3): The text of the Metz et al 2011 paper mentions that "the proposed method, registration was always performed on the complete 4D image and using a lung mask". The command line above does not call for a mask. Were different settings used in the Metz et al paper and was a single mask used for all images or was a 4D mask image used?
Question 4): Later work with ElastiX described in Delmon et al 2013, describes the use of direction dependent B-splines for relative motion between the lung and the thorax (http://elastix.bigr.nl/wiki/index.php/Par0016). This exercise requires a label image, that appears not to be readily available. Has any attempt been made to combine the methods in Par0012 and Par0016?
Thank you for your patience as I get my feet wet with your code.
Daniel R Einstein, PhD
Computational Biology & Bioinformatics
Pacific Northwest National Laboratory
Department of Mechanical Engineering
University of Washington, Seattle
Elastix mailing list
Elastix at bigr.nl<mailto:Elastix at bigr.nl>
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Elastix