@Tristan2Tzara_twitter : check out the odf_slicer in this example: https://dipy.org/documentation/1.0.0./examples_built/reconst_csd/
(for registration)
the right approach seems to be to mask the data before generating the scale space, but i'm not sure
i'm looking at this in the context of nipy/dipy#1969
this might be out of scope w.r.t. masking for image registration, but I notice that a lot of these registration functions / methods take 4 arguments at a minimum -- moving, fixed, moving_grid2world, fixed_grid2world -- and there are a lot of nearly duplicated code blocks which perform some operation on moving, then the same operation on fixed. Is there any appetite to simplifying this with some lightweight OOP? e.g. an object that has a data property, a mask property, and a grid2world property would save a lot of LOC in the registration codebase
this is looking pretty hairy to me:
I would say more than oop the plan is to simplify some things in the reg API. Not sure yet we can take out these parameters completely at first stage. You may feel this is hairy now but after a while you see that it is quite useful. Also do look into dipy.workflows.align there we have a higher level api. Do you have a suggestion for how to reduce the number of parameters? Be happy to suggest ideas. We are good listeners :)
@Garyfallidis I agree, those parameters are all essential, and it's good if the entry point to registration takes raw numpy arrays. I was thinking more about introducing some data structures (maybe something as simple as a
namedtuple that would be used internally by the registration / optimization routines
but I don't have any clear ideas right now
downsampling images after blurring should greatly reduce memory load + compute time for large images
(for reference, the images I am working with are ~10000 x 10000 pixels, and i have thousands of these images, so performance improvements matter a lot here)
@arokem @mvgolub thank you for the response. This is actually the file I have been trying to edit but have had trouble editing and troubleshooting. I have made some progress but do not have the expertise to do it efficiently. It looks like the last update was about 4 months ago. Is this still a work in progress?
Nope. No SfN booth
@smasjoodi : noisy data? What is your SNR? (https://dipy.org/documentation/1.0.0./examples_built/snr_in_cc/#example-snr-in-cc)
Hey, just a general question out of curiosity: What's causing the actual b-values to differ slightly from what I selected in the scanner console? e.g. b0 is actually not zero but around 26. or b=1000 is sometimes 999.34 or 1001.18 etc. Is this measured or calculated somehow from the gradient system? and does it improve the results significantly to use those actual values instead of the theoretical ones?
@Quantum_Neuron_twitter : I don't know the answer to this, but if no one here can provide an answer, I think that it would be a great question to ask on https://neurostars.org/
I do know that in some cases b-values also vary spatially
(empirical b-values, that is)
@arokem @RafaelNH @mvgolub I was investigating and it seems the post that was linked in a previous response to nipy/dipy#1744 has not been updated since May 7. Is there a dipy work in progress repository where newer files might be stored? Thanks again for the help
Hi @willi3by! @mvgolub did sime very nice progress on this code. We didn't push any update because we are testing it offline in deep. At the moment, we already have the same performance when comparing to the state of art of single shell algorithms in the recent literature, however we are still not happy with what single shell algorithms can give us. We are still detecting massive issues with this technique in general, and we are trying the best to overcome its major limitations. By the end of the year, we are planing: 1) to have our advances pushed in dipy; 2) share a manuscript reporting the details of our implementations and tests. Keep posted!
Hi @smasjoodi! Actually, considering that you have only 15 directions per shell, your maps looks good. DKI provides you a nice quantification of non-Gaussian diffusion (information that you cannot capture using techniques such as DTI), however its limitation is that is noisy, particularly in perpendicular directions where diffusivities are low (e.g. see chapter 3 of https://core.ac.uk/download/pdf/12428129.pdf). Radial and axial kurtosis can also have a noisier aspect because these estimates are not well define in crossing white matter regions and grey matter, such and axial and radial diffusivity (https://www.sciencedirect.com/science/article/pii/S1053811915000968?via%3Dihub). To improve DKI performance try to Denoise your data, and also Gibbs unring (e.g. see chapter 3 of https://www.repository.cam.ac.uk/handle/1810/281993) - code available in dipy (https://dipy.org/documentation/1.0.0./examples_built/denoise_mppca/#example-denoise-mppca) (https://dipy.org/documentation/1.0.0./examples_built/denoise_gibbs/#example-denoise-gibbs). Also the mean signal DKI might provide you an alternative and more precise non-gaussian diffusion characterisation (https://dipy.org/documentation/1.0.0./examples_built/reconst_msdki/#example-reconst-msdki).
any help appreciated!
Depends what you want to do. An affine transform may change the scale of your patient's brain. If you want to simply match the image matrices consider ResampleImage in ants; or convert3d and you may also want to check https://github.com › rordenlab › spmScripts › blob › nii_reslice_target. It is so good to have options! Best of luck!
@RafaelNH can you point me to examples of massive issues in single shell methods, or maybe not at this time? Just curious. Thanks and the best of luck!
@nabsabraham ; nice thing is you can do this in dipy !!! https://dipy.org/documentation/1.0.0./examples_built/affine_registration_3d/#example-affine-registration-3d excerpt: We can see that the images are far from aligned by drawing one on top of the other. The images don’t even have the same number of voxels, so in order to draw one on top of the other we need to resample the moving image on a grid of the same dimensions as the static image, we can do this by “transforming” the moving image using an identity transform
identity = np.eye(4)
affine_map = AffineMap(identity,
static.shape, static_grid2world,
moving.shape, moving_grid2world)
resampled = affine_map.transform(moving)
Every time I want to install dipy on my Anaconda it dipy0.16 is installed.
dipy conda-forge/linux-64::dipy-0.16.0-py27h4df700f_0
Hi @portokalh , thanks for the great answer!
I am planning to use a rigid body transformation as I don't want to deform anything just yet - however, I am working with dicoms and was wondering if there is any simple method to retrieve the affine voxel-to-patient matrix to do the registration
I am planning to use a rigid body transformation as I don't want to deform anything just yet - however, I am working with dicoms and was wondering if there is any simple method to retrieve the affine voxel-to-patient matrix to do the registration