users > Optical projection tomography and morphometry
Showing 1-14 of 14 posts
Mar 26, 2014 02:03 AM | Murat Maga
Optical projection tomography and morphometry
Dear all,
I have a large sample of mouse embryos (about 600) that are imaged in 3D using optical projection tomography (OPT). OPT is a low-cost and high-resolution alternative to mMRI for small objects (less than 2 cms) that can be chemically cleared and it uses UV spectrum.
My research question involves finding local and global differences in the head shape of these embryos between the control and alcohol exposed groups. So far I have been using landmark based generalized procrustes analysis to understand the shape. But for many different reason, it is far from adequate. I want to give voxel based morphometry a try but as I understand everything is more or less optimized for human MRI imaging.
How difficult would it be to build custom atlases from OPT datsets using CMTK? Is there anything in the underlying algorithms that require MRI modality? For example, intensities in the OPT does not correspond to any tissue type, in fact they vary depending on the age of the sample and how well it is cleared during the chemical processing.
Thanks,
Murat
I have a large sample of mouse embryos (about 600) that are imaged in 3D using optical projection tomography (OPT). OPT is a low-cost and high-resolution alternative to mMRI for small objects (less than 2 cms) that can be chemically cleared and it uses UV spectrum.
My research question involves finding local and global differences in the head shape of these embryos between the control and alcohol exposed groups. So far I have been using landmark based generalized procrustes analysis to understand the shape. But for many different reason, it is far from adequate. I want to give voxel based morphometry a try but as I understand everything is more or less optimized for human MRI imaging.
How difficult would it be to build custom atlases from OPT datsets using CMTK? Is there anything in the underlying algorithms that require MRI modality? For example, intensities in the OPT does not correspond to any tissue type, in fact they vary depending on the age of the sample and how well it is cleared during the chemical processing.
Thanks,
Murat
Mar 26, 2014 03:03 AM | Torsten Rohlfing
RE: Optical projection tomography and morphometry
Hi Murat -
There are actually at least three different techniques in CMTK that you can use to make a custom atlas from a set of images: groupwise nonrigid registration, iterative averaging, and averaging of registrations to a common reference.
None of these make any assumptions that would limit them to MRI. In fact, they have been extensively used for confocal microscopies of various insect brains.
The main question regarding your data is not whether a given tissue corresponds to a given image intensity, but whether there is some systematic relationship between the intensities of corresponding structures in different images. If that is not the case, then it is extremely unlikely that you would see your samples aligned in any meaningful way.
So while the exact intensity of any given structure or t issue is arbitrary and irrelevant, CMTK's tools would require, for example, that two distinct structures or tissues have distinct intensities, and that boundaries between different structures or tissues have some expression in the images, e.g., in the form of intensity gradients.
Ultimately, the proof is in the pudding... Just try some of CMTK's tools and see if the results are useful. If they are, hooray! If they are not, we have all learned something :-)
Feel free to use this forum to ask for specific suggestions, report problems, etc. If you want, maybe you could post a couple of example images (not the full stacks, just some relevant slices perhaps). That would make it much easier to say what techniques you might want to try.
Best,
Torsten
There are actually at least three different techniques in CMTK that you can use to make a custom atlas from a set of images: groupwise nonrigid registration, iterative averaging, and averaging of registrations to a common reference.
None of these make any assumptions that would limit them to MRI. In fact, they have been extensively used for confocal microscopies of various insect brains.
The main question regarding your data is not whether a given tissue corresponds to a given image intensity, but whether there is some systematic relationship between the intensities of corresponding structures in different images. If that is not the case, then it is extremely unlikely that you would see your samples aligned in any meaningful way.
So while the exact intensity of any given structure or t issue is arbitrary and irrelevant, CMTK's tools would require, for example, that two distinct structures or tissues have distinct intensities, and that boundaries between different structures or tissues have some expression in the images, e.g., in the form of intensity gradients.
Ultimately, the proof is in the pudding... Just try some of CMTK's tools and see if the results are useful. If they are, hooray! If they are not, we have all learned something :-)
Feel free to use this forum to ask for specific suggestions, report problems, etc. If you want, maybe you could post a couple of example images (not the full stacks, just some relevant slices perhaps). That would make it much easier to say what techniques you might want to try.
Best,
Torsten
Mar 26, 2014 04:03 AM | Murat Maga
RE: Optical projection tomography and morphometry
Hi Torsten,
Thank you for your response. I attached a few cross-section to give you an idea about what the data looks like. I am not sure if I follow your statement about "systematic relationship between the intensities of corresponding structures in different images". Can you comment whether those sample images fit to this criteria
Since I am interested in the entire head, I thought it would be fairly straighforward to segment it via thresholding. But I am guessing that's not going to be sufficient for non-rigid registration.
Thank you for your response. I attached a few cross-section to give you an idea about what the data looks like. I am not sure if I follow your statement about "systematic relationship between the intensities of corresponding structures in different images". Can you comment whether those sample images fit to this criteria
Since I am interested in the entire head, I thought it would be fairly straighforward to segment it via thresholding. But I am guessing that's not going to be sufficient for non-rigid registration.
Mar 26, 2014 09:03 PM | Torsten Rohlfing
RE: Optical projection tomography and morphometry
Hi Murat -
I looked at your images, and I see no reason why you shouldn't be able to create a custom atlas with CMTK's tools.
Whether the result will be useful will of course depend on your expectations. In general, CMTK's alignment tools (like most others that I am aware of) compute spatially continuous, somewhat smooth coordinate mappings. So if you have topological differences between your images (e.g., two structures of some type in one image, but only one in another) then you will not be able to get these reconciled. Also, CMTK may (will) not be able to model every last detail of variation between complex-shaped structures.
But in general, I'd recommend you give this a shot. My suggestion would be to start with the "iterative_shape_averaging" script, which is losely based on the algorithm in this paper: http://dx.doi.org/10.1109/MMBIA.2001.991...
You may not want to run this on your entire dataset though (600 you said?) Greg Jefferis for his 2007 Cell paper used the algorithm on a smaller sample (I believe 20?) of particularly well-imaged cases to make a template, then used that template as the reference to align all cases from his complete collection to that via pairwise, image-to-image registration.
As for my remark about the "systematic relationships" - never mind ;) Sometimes registration has something of "black magic" to it, where it's not entirely clear all the time if something works, why it does or does not, and what exactly to do to make it work. ;)
Best,
Torsten
PS: Greg is maintaining a high-level Perl script, "munger", that comes bundled with CMTK. I never quite understood what exactly it does, but it might be useful for you :)
I looked at your images, and I see no reason why you shouldn't be able to create a custom atlas with CMTK's tools.
Whether the result will be useful will of course depend on your expectations. In general, CMTK's alignment tools (like most others that I am aware of) compute spatially continuous, somewhat smooth coordinate mappings. So if you have topological differences between your images (e.g., two structures of some type in one image, but only one in another) then you will not be able to get these reconciled. Also, CMTK may (will) not be able to model every last detail of variation between complex-shaped structures.
But in general, I'd recommend you give this a shot. My suggestion would be to start with the "iterative_shape_averaging" script, which is losely based on the algorithm in this paper: http://dx.doi.org/10.1109/MMBIA.2001.991...
You may not want to run this on your entire dataset though (600 you said?) Greg Jefferis for his 2007 Cell paper used the algorithm on a smaller sample (I believe 20?) of particularly well-imaged cases to make a template, then used that template as the reference to align all cases from his complete collection to that via pairwise, image-to-image registration.
As for my remark about the "systematic relationships" - never mind ;) Sometimes registration has something of "black magic" to it, where it's not entirely clear all the time if something works, why it does or does not, and what exactly to do to make it work. ;)
Best,
Torsten
PS: Greg is maintaining a high-level Perl script, "munger", that comes bundled with CMTK. I never quite understood what exactly it does, but it might be useful for you :)
Mar 26, 2014 09:03 PM | Murat Maga
RE: Optical projection tomography and morphometry
Hi Torsten,
Yes, I have (or will) about 600 samples, but this is a time series. I have 8 developmental time points, and two groups per time point (exposure and ctrl), making 16 total groups, so give or take about 30-35 samples in each sub category.
One thing I should mention is that the morphological change from 1st time point to the last is huge. So I guess, I will have to do separate ones for each time point. But then for a given time point, should I build the atlas based on the controls only, or based on both groups together?
I never did voxel based morphometry before. At this point I am trying to understand how to design the workflow.
Thank you again for your time.
M
Yes, I have (or will) about 600 samples, but this is a time series. I have 8 developmental time points, and two groups per time point (exposure and ctrl), making 16 total groups, so give or take about 30-35 samples in each sub category.
One thing I should mention is that the morphological change from 1st time point to the last is huge. So I guess, I will have to do separate ones for each time point. But then for a given time point, should I build the atlas based on the controls only, or based on both groups together?
I never did voxel based morphometry before. At this point I am trying to understand how to design the workflow.
Thank you again for your time.
M
Mar 26, 2014 11:03 PM | Greg Jefferis
RE: Optical projection tomography and morphometry
I would start by trying to register one time point/genotype against
a single good brain. Make sure you have good even staining that the
images have a consistent field of view (presumably whole brain
within stack) and are rotated in the same way.
You will likely need to optimise the parameters. To get started you can either use Torsten's semi-automated command line tools rgistrationx/warpx and the like or try and use the simple GUI wrappers that some have found useful:
http://flybrain.mrc-lmb.cam.ac.uk/dokuwi...
It will be something of a matter of chance whether your spatial resolution / number of voxels means that the parameters we have used for fly registrations make sense. You can try reading this to see if it helps:
http://flybrain.mrc-lmb.cam.ac.uk/dokuwi...
but the best is probably a suggestion from Torsten. One point, in the early days we wasted quite a bit of time registering at too high a final resolution (in terms of the non-rigid registration step size being very small compared with the voxel size).
Good luck!
Greg.
You will likely need to optimise the parameters. To get started you can either use Torsten's semi-automated command line tools rgistrationx/warpx and the like or try and use the simple GUI wrappers that some have found useful:
http://flybrain.mrc-lmb.cam.ac.uk/dokuwi...
It will be something of a matter of chance whether your spatial resolution / number of voxels means that the parameters we have used for fly registrations make sense. You can try reading this to see if it helps:
http://flybrain.mrc-lmb.cam.ac.uk/dokuwi...
but the best is probably a suggestion from Torsten. One point, in the early days we wasted quite a bit of time registering at too high a final resolution (in terms of the non-rigid registration step size being very small compared with the voxel size).
Good luck!
Greg.
Mar 26, 2014 11:03 PM | Greg Jefferis
RE: Optical projection tomography and morphometry
PS what I was referring to re semi-automated command line tools was
this option of
reformatx
--auto-multi-levels
Automatic optimization and resolution parameter generation for
levels
[Default: 0]
But I had forgotten that it does not exist for warpx.
Best,
Greg.
reformatx
--auto-multi-levels
Automatic optimization and resolution parameter generation for
levels
[Default: 0]
But I had forgotten that it does not exist for warpx.
Best,
Greg.
Mar 27, 2014 02:03 AM | Torsten Rohlfing
RE: Optical projection tomography and morphometry
Soooo.... *voxel*-based morphometry, huh? Well, thing is, I
don't think I've done that, ever... so take any advice from me on
the subject with a grain of salt.
That said - I believe the general consensus of the VBM community goes towards creating an atlas from both controls and exposed cases. so as to avoid bias in favor of one of the groups simply by virtue of being more similar to the atlas.
As for your time series - my question is, can you tell precisely for each sample what time point it should go with? If there is uncertainty, you might be better served by a 4D (3D+t) atlas such as the one we've described here: http://dx.doi.org/10.1007/978-3-642-0249...
The tools are in CMTK, but that's where stuff gets really complicated... so maybe not a good start.
Anyway, one important thing you need to decide - what is it you are actually going to try and quantify after atlas construction and presumably alignment of your samples to the atlas(es)? In human MRI VBM, people usually quantify local tissue (CSF/GM/WM) volume, after segmenting each sample, reformatting into atlas space, and smoothing the hell out of it to compensate for the typically lousy registration accuracy ;)
Best,
Torsten
That said - I believe the general consensus of the VBM community goes towards creating an atlas from both controls and exposed cases. so as to avoid bias in favor of one of the groups simply by virtue of being more similar to the atlas.
As for your time series - my question is, can you tell precisely for each sample what time point it should go with? If there is uncertainty, you might be better served by a 4D (3D+t) atlas such as the one we've described here: http://dx.doi.org/10.1007/978-3-642-0249...
The tools are in CMTK, but that's where stuff gets really complicated... so maybe not a good start.
Anyway, one important thing you need to decide - what is it you are actually going to try and quantify after atlas construction and presumably alignment of your samples to the atlas(es)? In human MRI VBM, people usually quantify local tissue (CSF/GM/WM) volume, after segmenting each sample, reformatting into atlas space, and smoothing the hell out of it to compensate for the typically lousy registration accuracy ;)
Best,
Torsten
Mar 27, 2014 03:03 AM | Murat Maga
RE: Optical projection tomography and morphometry
This is a timed breeding experiment, I do know the exact times
(within about 6hrs) of each sample. I guess I do not have to worry
about the 4D atlas.
Since the landmarks are not performing well, and I have the volumes, VBM seemed the logical thing to try. How much of it I can do we shall see. I am naively thinking to use the whole skull as single structure to quantify the shape differences. But if you have other suggestion or if my thinking is flawed, do let me know. How would you have done it?
M
Since the landmarks are not performing well, and I have the volumes, VBM seemed the logical thing to try. How much of it I can do we shall see. I am naively thinking to use the whole skull as single structure to quantify the shape differences. But if you have other suggestion or if my thinking is flawed, do let me know. How would you have done it?
M
Mar 27, 2014 04:03 PM | Torsten Rohlfing
RE: Optical projection tomography and morphometry
Question - how are you planning to quantify "shape differences"
using a VBM-type approach? The thing about VBM is that it gives you
some local measure of which group has how much of what at a given
location (and within a small region around it). But shape, at least
as I understand it, is more of a global property.
To me, that calls for a whole different class of algorithms, such as m-reps (out of UNC) or spherical harmonics. None of these CMTK would be any help with though, but I believe there are a number of projects on NITRC that cover these.
TR
To me, that calls for a whole different class of algorithms, such as m-reps (out of UNC) or spherical harmonics. None of these CMTK would be any help with though, but I believe there are a number of projects on NITRC that cover these.
TR
Mar 27, 2014 06:03 PM | Murat Maga
RE: Optical projection tomography and morphometry
Hi Torsten,
It is possible that I misunderstood what VBM does. I am looking for ways to see localized shape differences in the face. It is possible that (and in fact is known) the eyes of the individual affected individual would be small. I can only put 4 landmarks around the eye to approximate the shape. I was hoping a technique like VBM would display the difference visually and put a probability based on provided sample.
The idea is to find other regions with high likelihood of difference between affected and controls without restricting ourselves to landmarks. We already know something is happening with the eyes, overall skullsize, lips (shapes of which can be reasonably captured by landmark). Question is what's happening with the rest. Are we losing information because we can't identify landmarks? With landmarks it is not easy to see what might be happening on cheek, lateral walls of nose, forehead all of which areas devoid of prominent anatomical landmarks.
Secondary reason is my desire to move away having to do this analysis by manually collecting thousands of landmarks. I thought VBM can provide all of the above.
And finally thank you very much your responses. It is being very useful for me and I appreciate it.
M
It is possible that I misunderstood what VBM does. I am looking for ways to see localized shape differences in the face. It is possible that (and in fact is known) the eyes of the individual affected individual would be small. I can only put 4 landmarks around the eye to approximate the shape. I was hoping a technique like VBM would display the difference visually and put a probability based on provided sample.
The idea is to find other regions with high likelihood of difference between affected and controls without restricting ourselves to landmarks. We already know something is happening with the eyes, overall skullsize, lips (shapes of which can be reasonably captured by landmark). Question is what's happening with the rest. Are we losing information because we can't identify landmarks? With landmarks it is not easy to see what might be happening on cheek, lateral walls of nose, forehead all of which areas devoid of prominent anatomical landmarks.
Secondary reason is my desire to move away having to do this analysis by manually collecting thousands of landmarks. I thought VBM can provide all of the above.
And finally thank you very much your responses. It is being very useful for me and I appreciate it.
M
Mar 27, 2014 08:03 PM | Torsten Rohlfing
RE: Optical projection tomography and morphometry
Murat:
I think you may want to look into Deformation-Based Morphometry (DBM) rather than VBM.
Here's why - let's say you have a perfect registration algorithm to align your samples to the template (and of course we all constantly strive for that perfect algorithm ;)) That means, after alignment all your samples will essentially look the same - just like your template/atlas. So doing a voxelwise comparison won't tell you anything.
That is because any difference in the shape/appearance of your samples has been moved out of the image domain and is now encoded solely in the deformation that maps your sample to your atlas. So then DBM basically looks at these mappings and compares those (rather than the image data itself, as VBM would).
Another option would be this - make separate templates for your two groups, align the templates using nonrigid registration, and derive conclusions from that mapping. You might find this paper an interesting read on the two options:
T. Rohlfing, A. Pfefferbaum, E. V. Sullivan, and C. R. Maurer, Jr., “Information fusion in biomedical image analysis: Combination of data vs. combination of interpretations,” in Information Processing in Medical Imaging, G. E. Christensen and M. Sonka, Eds., Berlin, Germany, 2005, vol. 3565 of Lecture Notes in Computer Science, pp. 150-161, Springer-Verlag, 19th International Conference, IPMI 2005, Glenwood Springs, CO, USA, July 10-15, 2005.
Admittedly, though, to take full advantage of your entire data set, you'd probably be better off doing a "standard" DBM with a single template, aligning each sample to it, then analyzing the deformation fields that result.
Best,
Torsten
I think you may want to look into Deformation-Based Morphometry (DBM) rather than VBM.
Here's why - let's say you have a perfect registration algorithm to align your samples to the template (and of course we all constantly strive for that perfect algorithm ;)) That means, after alignment all your samples will essentially look the same - just like your template/atlas. So doing a voxelwise comparison won't tell you anything.
That is because any difference in the shape/appearance of your samples has been moved out of the image domain and is now encoded solely in the deformation that maps your sample to your atlas. So then DBM basically looks at these mappings and compares those (rather than the image data itself, as VBM would).
Another option would be this - make separate templates for your two groups, align the templates using nonrigid registration, and derive conclusions from that mapping. You might find this paper an interesting read on the two options:
T. Rohlfing, A. Pfefferbaum, E. V. Sullivan, and C. R. Maurer, Jr., “Information fusion in biomedical image analysis: Combination of data vs. combination of interpretations,” in Information Processing in Medical Imaging, G. E. Christensen and M. Sonka, Eds., Berlin, Germany, 2005, vol. 3565 of Lecture Notes in Computer Science, pp. 150-161, Springer-Verlag, 19th International Conference, IPMI 2005, Glenwood Springs, CO, USA, July 10-15, 2005.
Admittedly, though, to take full advantage of your entire data set, you'd probably be better off doing a "standard" DBM with a single template, aligning each sample to it, then analyzing the deformation fields that result.
Best,
Torsten
Mar 27, 2014 08:03 PM | Torsten Rohlfing
RE: Optical projection tomography and morphometry
By the way - do keep those landmarks you've already collected;
those will make for a great gold standard to validate registration
accuracy. (Which, in turn, will greatly facilitate your attempts to
eventually get your results published ;))
TR
TR
Mar 27, 2014 08:03 PM | Torsten Rohlfing
RE: Optical projection tomography and morphometry
One more, somewhat in response to myself:
Having suggested DBM instead of VBM, I should mention that CMTK's abilities in that area are somewhat constrained.
Since all of CMTK's nonrigid registration algorithms rely on the B-spline free-form deformation transformation model, their resolution is limited by their smoothness conditions. CMTK is also not terribly good at computing invertible (i.e., non-zero Jacobian) transformations, let alone compute the actual inverse of a transformation along with the transformation itself.
If you find yourself limited by these issues, then I can recommend Brian Avants' ANTs toolkit, specifically the "symmetric normalization" (SyN) algorithm, which computes unbiased, diffeomorphic transformations and their inverses.
That said, these problems mostly refer to the transformations between your ultimate template and each sample. For actually creating a template, CMTK should be fine.
Best,
Torsten
Having suggested DBM instead of VBM, I should mention that CMTK's abilities in that area are somewhat constrained.
Since all of CMTK's nonrigid registration algorithms rely on the B-spline free-form deformation transformation model, their resolution is limited by their smoothness conditions. CMTK is also not terribly good at computing invertible (i.e., non-zero Jacobian) transformations, let alone compute the actual inverse of a transformation along with the transformation itself.
If you find yourself limited by these issues, then I can recommend Brian Avants' ANTs toolkit, specifically the "symmetric normalization" (SyN) algorithm, which computes unbiased, diffeomorphic transformations and their inverses.
That said, these problems mostly refer to the transformations between your ultimate template and each sample. For actually creating a template, CMTK should be fine.
Best,
Torsten