DTI and DKI processing
Hello Dipy Developers, My name is Suren. I am doing group-wise analysis of Alzheimer’s disease and healthy controls normal brain. I could successfully run DTI and DKI analysis. I did all other preprocessing steps: 1. Denoising (dwidenoise from MRtrix3) 2. Gibb’s artifact removal (from MRtrix3) 3. Eddy current correction and motion artifact removal (affine registration) Yet, I am wondering if we need data normalization and data regularization before performing DTI and DKI analysis. Also, I am wondering if we need to remove outliers before conducting both. I am using 3 shell dMRI data acquired in 9.4 Tesla MRI Scanner. I am waiting for your favorable response. Thank you very much. Best Regards, Suren Indiana University
Hi Suren, I am not 100% sure what "data normalization and regularization" means in this context. One thing you might consider for dealing with outliers is to use the RESTORE method for DTI, which robustifies the model fit against outlying observations. Cheers, Ariel On Wed, Nov 10, 2021 at 10:26 AM Maharjan, Surendra <smaharj@iu.edu> wrote:
Hi Ariel, The data normalization here I would like to know: Can we normalize the DWI data in the range 0 to 1 before fitting? Thank you very much. I will work on RESTORE DTI fit. Regularization here I mean: L1 (Lasso) and L2 (Ridge) regularization. Do we have similar robust fit like RESTORE for DKI? Thank you, Ariel. Your comments always saved me. Best Regards, Suren From: Ariel Rokem <arokem@uw.edu> Date: Wednesday, November 10, 2021 at 1:34 PM To: Maharjan, Surendra <smaharj@iu.edu> Cc: dipy@python.org <dipy@python.org> Subject: [External] Re: [DIPY] DTI and DKI processing This message was sent from a non-IU address. Please exercise caution when clicking links or opening attachments from external sources. Hi Suren, I am not 100% sure what "data normalization and regularization" means in this context. One thing you might consider for dealing with outliers is to use the RESTORE method for DTI, which robustifies the model fit against outlying observations. Cheers, Ariel On Wed, Nov 10, 2021 at 10:26 AM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hello Dipy Developers, My name is Suren. I am doing group-wise analysis of Alzheimer’s disease and healthy controls normal brain. I could successfully run DTI and DKI analysis. I did all other preprocessing steps: 1. Denoising (dwidenoise from MRtrix3) 2. Gibb’s artifact removal (from MRtrix3) 3. Eddy current correction and motion artifact removal (affine registration) Yet, I am wondering if we need data normalization and data regularization before performing DTI and DKI analysis. Also, I am wondering if we need to remove outliers before conducting both. I am using 3 shell dMRI data acquired in 9.4 Tesla MRI Scanner. I am waiting for your favorable response. Thank you very much. Best Regards, Suren Indiana University _______________________________________________ DIPY mailing list -- dipy@python.org<mailto:dipy@python.org> To unsubscribe send an email to dipy-leave@python.org<mailto:dipy-leave@python.org> https://mail.python.org/mailman3/lists/dipy.python.org/ Member address: arokem@gmail.com<mailto:arokem@gmail.com>
Hi Suren, On Wed, Nov 10, 2021 at 10:57 AM Maharjan, Surendra <smaharj@iu.edu> wrote:
You can, but that kind of normalization is already baked into DTI and DKI fitting, so it won't change your model fits.
I don't think it makes sense to use these for DTI and/or DKI, which have only a few regressors that are (I think) orthogonal by construction.
Do we have similar robust fit like RESTORE for DKI?
Yeah - I think that you can use RESTORE in DKI as well (initialize the model with `fit_method="RESTORE"` and with a `sigma` value). The more thorough solution for that would be to implement Tax et al.'s REKINDLE method (https://onlinelibrary.wiley.com/doi/full/10.1002/mrm.25165), but we don't have that. Cheers, Ariel
Dear Dipy fellas, My student would like to use many data from the ADNI dataset, but we noticed a lot of heterogeneity. Even taking the data from the same type of scanner and same field (we checked both ADNI1 ADNI2 and ADNI3), there is1. Different TR and TE 2. Different number of gradients We found the COMBAT tool https://github.com/Jfortin1/ComBatHarmonization, which according to my students is a nightmare and also does not solve the issue of different number of gradients. How can we address this?I know a solution would be to drop the ADNI dataset and use something else, but there are other data in this dataset which are not available in other datasets. I know Bramsh Chandio worked on the PPMI dataset which I imagine has the same level of messiness... or not? Best,Alex Prof. Dr. Alessandro Crimi Research Group Leader https://bam.sano.science Centre for Computational Medicine Czarnowiejska 36, building C5. 30-072 Kraków, Poland Phone: +48 575 453 005 www.sano.science ----------------------------------------------- Visiting Lecturer African Institute for Mathematical Sciences www.aims.edu.gh --------------------------------------------------------- https://twitter.com/Dr_Alex_Crimi
Hi Alex, On Thu, Nov 11, 2021 at 5:31 AM Alex Crimi via DIPY <dipy@python.org> wrote:
Dear Dipy fellas,
Hi Alex, Thanks for raising this! It's a tough issue.
I'll let Bramsh chime in about PPMT, but I will mention that we've used COMBAT for harmonization at the level of the diffusion features / tract profiles. We've found it to be rather effective in harmonizing across sites in the HBN study, despite some frustrating differences across sites, though not as much variability in the acquisition protocols (see supplemental S12 here: https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.10091...). This might also address different acquisition parameters to some degree, assuming you have enough subjects in each acquisition protocol. We've found this package to be slightly more user-friendly than the implementation you pointed out: https://github.com/Warvito/neurocombat_sklearn, because it follows the sklearn API. Cheers, Ariel
Hi,thanks for the links. So you confirm you have been able to harmonize given different number of gradient directions?I cannot find this information in your paper.Best,Alex On Thursday, November 11, 2021, 06:34:16 PM GMT+1, Ariel Rokem <arokem@uw.edu> wrote: Hi Alex, On Thu, Nov 11, 2021 at 5:31 AM Alex Crimi via DIPY <dipy@python.org> wrote: Dear Dipy fellas, Hi Alex, Thanks for raising this! It's a tough issue. My student would like to use many data from the ADNI dataset, but we noticed a lot of heterogeneity. Even taking the data from the same type of scanner and same field (we checked both ADNI1 ADNI2 and ADNI3), there is1. Different TR and TE 2. Different number of gradients We found the COMBAT tool https://github.com/Jfortin1/ComBatHarmonization, which according to my students is a nightmare and also does not solve the issue of different number of gradients. How can we address this? I'll let Bramsh chime in about PPMT, but I will mention that we've used COMBAT for harmonization at the level of the diffusion features / tract profiles. We've found it to be rather effective in harmonizing across sites in the HBN study, despite some frustrating differences across sites, though not as much variability in the acquisition protocols (see supplemental S12 here: https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.10091...). This might also address different acquisition parameters to some degree, assuming you have enough subjects in each acquisition protocol. We've found this package to be slightly more user-friendly than the implementation you pointed out: https://github.com/Warvito/neurocombat_sklearn, because it follows the sklearn API. Cheers, Ariel I know a solution would be to drop the ADNI dataset and use something else, but there are other data in this dataset which are not available in other datasets. I know Bramsh Chandio worked on the PPMI dataset which I imagine has the same level of messiness... or not? Best,Alex Prof. Dr. Alessandro Crimi Research Group Leader https://bam.sano.science Centre for Computational Medicine Czarnowiejska 36, building C5. 30-072 Kraków, Poland Phone: +48 575 453 005 www.sano.science ----------------------------------------------- Visiting Lecturer African Institute for Mathematical Sciences www.aims.edu.gh --------------------------------------------------------- https://twitter.com/Dr_Alex_Crimi _______________________________________________ DIPY mailing list -- dipy@python.org To unsubscribe send an email to dipy-leave@python.org https://mail.python.org/mailman3/lists/dipy.python.org/ Member address: arokem@gmail.com _______________________________________________ DIPY mailing list -- dipy@python.org To unsubscribe send an email to dipy-leave@python.org https://mail.python.org/mailman3/lists/dipy.python.org/ Member address: alex.crimi@yahoo.com
Hi Alex, On Thu, Nov 11, 2021 at 9:48 AM Alex Crimi <alex.crimi@yahoo.com> wrote:
No - sorry - the acquisitions in HBN have the same number of gradient directions in all of the sites, so much less variable than what you are facing with ADNI. Cheers, Ariel
Hi Alex, Ariel, Just to add another citation, ADNI3 protocol analyses using ROIs+diffusion features/indices and ComBat can be found here: https://internal-journal.frontiersin.org/articles/10.3389/fninf.2019.00002/f... Supplementary figures 1 and 2 are perhaps the most relevant, which seem to show that ComBat is removing differences for linear associations. Another option though would be to harmonize in the local model parameter space, instead of in the original spaces. Hope this helps, Daniel Moyer On Thu, Nov 11, 2021 at 12:34 PM Ariel Rokem <arokem@uw.edu> wrote:
Hi everyone, Check this out: https://github.com/pnlbwh/dMRIharmonization Best, Suheyla Cetin-Karayumak, PhD Assistant Professor, Psychiatry Neuroimaging Laboratory, Brigham and Women's Hospital, Harvard Medical School, Boston. skarayumak@bwh.harvard.edu ________________________________ From: Ariel Rokem <arokem@uw.edu> Sent: Thursday, November 11, 2021 12:14 PM To: Alex Crimi <alex.crimi@yahoo.com> Cc: dipy@python.org <dipy@python.org>; bqchandi@iu.edu <bqchandi@iu.edu> Subject: [DIPY] Re: Hamonizing ADNI for some Dipy stuff External Email - Use Caution Hi Alex, On Thu, Nov 11, 2021 at 5:31 AM Alex Crimi via DIPY <dipy@python.org<mailto:dipy@python.org>> wrote: Dear Dipy fellas, Hi Alex, Thanks for raising this! It's a tough issue. My student would like to use many data from the ADNI dataset, but we noticed a lot of heterogeneity. Even taking the data from the same type of scanner and same field (we checked both ADNI1 ADNI2 and ADNI3), there is 1. Different TR and TE 2. Different number of gradients We found the COMBAT tool https://github.com/Jfortin1/ComBatHarmonization<https://secure-web.cisco.com/1qYv-oUqGK6oTHJSqWwxscxc52VJxb2jOkmS7yFdecQIJNuuCiSoREhwuTKVlNtzhG3hfvoR5UqiopXrRtqK4Ob68mBw6MpSN4ufv44D8ldLMFslolNQiKN8CMdEWxc5khutYaFynJdS-A-1o-ZdNYfx9wUiBK-fpoEJ1u0pzslcEZdOfSaQddDHw4QMjtDizSCR_7Blx0_rm5djrduVisSGqW69fw74DTAkwngTNm_O-SdladRDKFsczre-pB2EZY6U1gEWrPHZ7RjbKNkhCpVt_XwIDt62eT9Y6A1UX_wIVwgvnkehhSJjI7YRZazGU/https%3A%2F%2Fgithub.com%2FJfortin1%2FComBatHarmonization>, which according to my students is a nightmare and also does not solve the issue of different number of gradients. How can we address this? I'll let Bramsh chime in about PPMT, but I will mention that we've used COMBAT for harmonization at the level of the diffusion features / tract profiles. We've found it to be rather effective in harmonizing across sites in the HBN study, despite some frustrating differences across sites, though not as much variability in the acquisition protocols (see supplemental S12 here: https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1009136<https://secure-web.cisco.com/1yOHjaWE-bwF448WN4jS75U6VClF1g3YxyG8taPUdf1PEZhgDgBXzcizpeDg1yVw5D1eTNs3YyrxBAsSjNbAa9VG8o6qgevdkoqitclMRCd70mDq2bk_fjkkzOud2aiKzkQGLAp_Eyhxq58cCMktJOh8ba0YVN-k0dhLeoW4AB5wpIVLky_bW95PolWjC2QZlt72Ioqmp5Pu_PxBG8bbG6lqwZi5ze1apkvDmI8b7wbGYGapgfzivr-SMvyLStrvmrC517WXYWT5oFQS1J8SZ2AxBjdPi09Kv-pl-UXNh-HoW8NX9BZWypqXybkA-PZp8/https%3A%2F%2Fjournals.plos.org%2Fploscompbiol%2Farticle%3Fid%3D10.1371%2Fjournal.pcbi.1009136>). This might also address different acquisition parameters to some degree, assuming you have enough subjects in each acquisition protocol. We've found this package to be slightly more user-friendly than the implementation you pointed out: https://github.com/Warvito/neurocombat_sklearn<https://secure-web.cisco.com/1gL9DmOOTnMNnyUG6j3HgqAJ_n3YvoN6VZWCe8tE1u2_-VEnrI1s9SGW5GraEY0bT_sjIaxQ9Kku8suxXoi7oXO0D_613QwD4Ldov3MHLYpEGJ99tcL9zGh0pbEwJIZNrl7R5NV3P4u3MOTRzF8k-dxkX5l3R0N1WHsG3XqpqjuvYr02jitrFO_j1WI9FCxT0pBxrZ7yGGQipG_jSeVOunSr9PvSFtclU8F4wy4pSKzBYga_zGGdVnAfejK55xxRNQyl_om1Cs73oLzfs2enGhL2RR2-Fzd7PjKrS6K-RDvdlunflqraBg09PTWQd_Q2V/https%3A%2F%2Fgithub.com%2FWarvito%2Fneurocombat_sklearn>, because it follows the sklearn API. Cheers, Ariel I know a solution would be to drop the ADNI dataset and use something else, but there are other data in this dataset which are not available in other datasets. I know Bramsh Chandio worked on the PPMI dataset which I imagine has the same level of messiness... or not? Best, Alex Prof. Dr. Alessandro Crimi Research Group Leader https://bam.sano.science<https://secure-web.cisco.com/1P37o1AQpD282af3eR09BoxrOLJ9c8TRIL8nF3QyjlWf_80xB1hT3dp3ioKZmBx9gjKBq3JHpBLRLIhiCpdlBWLpLNX0N9hQdN80F3Z-7q5WL33BgjOXno6AaPjS9zmmjOV3ERdkpl1iJnf1qTEtjJMtVU0fKFL_7OdGVdBRx1RCcrGDGe4KUNGO844TliQ2JJJfaZIOv8iqtIBAdLt1nvbPj6t61zChSSNGMEH2BMNph0zjtd6EGbQl2nSJbxlw_RLBUdNigSLh6uOtFyo5OTnW3I26NqdywNVMhsmy3GXF6M9tOfNZQNjBO8rcuxRTM/https%3A%2F%2Fbam.sano.science> [X] Centre for Computational Medicine Czarnowiejska 36, building C5. 30-072 Kraków, Poland Phone: +48 575 453 005 www.sano.science<http://secure-web.cisco.com/1eGjBi4ybQ-RwvzMnhpzMGzXIH3QHrd4acViwk311yDFFOWk...> ----------------------------------------------- Visiting Lecturer African Institute for Mathematical Sciences www.aims.edu.gh<http://secure-web.cisco.com/1rMVeXDkP0BFhRBHJTIklU3_iREWl_6EvFl4ifvYOjSwD7k7...> --------------------------------------------------------- https://twitter.com/Dr_Alex_Crimi<https://secure-web.cisco.com/11JPLwI8c7Da7TrruKifMgFOOC4GbBFxJA1QF3zSjJlTQKcarnuC6-tmQMuUlNqBgE2UQ9n8eE0AoXnjaLIzM2LUYptrmEDAiNi4d0P46sv0IEeJX3gWe6OLurfH20C0vwZk7Y_myAkRewafeUnXKBcShaATXuIjq_4hh4jqFrlU0VDMdFAf8wb4Nmkdpst1ytEharGordws3_xleOA-ECc2dOKJYXctkn-u_iavFBCO5XQ7_i4lGMySfbcfZw7zhUFT382Ce7af5aVIt4WBHZtLjFmOz3OyROnRlsqSuHpyqxlJCm42ng5YOhRdPa0MR/https%3A%2F%2Ftwitter.com%2FDr_Alex_Crimi> _______________________________________________ DIPY mailing list -- dipy@python.org<mailto:dipy@python.org> To unsubscribe send an email to dipy-leave@python.org<mailto:dipy-leave@python.org> https://mail.python.org/mailman3/lists/dipy.python.org/<https://secure-web.cisco.com/1SPfzZRIgTfgpftXLS6ppjvH5VKfjmmP7WUkFeHWx5vEvwJOIt0f0IbIVj1BhTnwxXyGV1F3gn-HbgrL1dKC3JM3hsfwnH7lpCAYS-bF4wfLQVBApGk7NMuy3lsHbOQYBodcwJkImTdNNpdUSnoPU82P8OUYw_YI7VyL900cygT0ER7NiU1ZrJRFub11zINpIvZH5zqa4CAm3Z__WpJaW2swTDM2TmnxiZJXIaUpCAIxhWS9mQmpPHK6WuPpt_MH4GjjQuTEW_WNYu1T4OaYzk-05B-EsxGsJGYkxJPsGt5nRiG5g8m4-03cjk7eMG9xm/https%3A%2F%2Fmail.python.org%2Fmailman3%2Flists%2Fdipy.python.org%2F> Member address: arokem@gmail.com<mailto:arokem@gmail.com> The information in this e-mail is intended only for the person to whom it is addressed. If you believe this e-mail was sent to you in error and the e-mail contains patient information, please contact the Mass General Brigham Compliance HelpLine at http://www.massgeneralbrigham.org/complianceline . If the e-mail was sent to you in error but does not contain patient information, please contact the sender and properly dispose of the e-mail. Please note that this e-mail is not secure (encrypted). If you do not wish to continue communication over unencrypted e-mail, please notify the sender of this message immediately. Continuing to send or respond to e-mail after receiving this message means you understand and accept this risk and wish to continue to communicate over unencrypted e-mail.
Hi everyone, Also check the latest work on harmonization of diffusion MRI on ADNI using COMBAT, COMBAT-GAM and COVBAT by Sophia Thomopoulos. It will be presented at SIPAIM 2021 (https://sipaim.org): https://www.medrxiv.org/content/10.1101/2021.10.04.21263994v1.full.pdf Cheers, Julio El jue, 11 nov 2021 a las 10:42, Cetin Karayumak, Suheyla (< SKARAYUMAK@bwh.harvard.edu>) escribió:
Hello, ADNI questions are welcomed at http://adni.loni.usc.edu/support/experts-knowledge-base/ask-experts/ , where questions are forwarded to the appropriate experts according to their category. We don't expect other forums to provide answers for us, although dipy's has done a fine job. Data harmonization is an active area of research, but as noted in Sophia Thomopoulos' preprint ComBat does a good job for atlas region values, and Daniel Moyer's papers would be a good start for the tougher problem of making the images themselves more directly comparable. One approach that I don't think has been mentioned in this thread is to extract a homogeneous subset from the ADNI data. After all, the reason ADNI has heterogeneity is that it spans multiple scanners, and actually, multiple generations of scanners. If we had stuck to a homogeneous set of scanners the resulting data set would have been about the same size as pulling a homogeneous set out of ADNI. In practice when selecting a set it is best to know which differences can be ignored, which can be harmonized reasonably easily, and which should be avoided. ADNI has had 3 stages so far, each lasting about 5 years. ADNI1: Did not have dMRI. I was not involved with ADNI until much later, but my understanding is that any "ADNI1 DTIs" you might find were experimental and a standard protocol had not been settled on yet. ADNI2: Actually was relatively homogeneous, because it was restricted to GE, with 5 b = 0 + 41 b = 1000, 2.7mm resolution. You might find a few 4 + 41s - I would treat them as the same. You might also find a few with 60 b = 1000. Those were *not* the standard ADNI2 protocol, and typically have a matching standard DTI for the same participant and date. i.e. they're not for people wanting a homogeneous set, they're for people *wanting* to try out their harmonization method. (Even though at the time the main motivation was likely testing out an enhanced number of directions.) ADNI3: Deliberately broke with ADNI2, in order to keep up with advances in the field, and to increase the number of dMRIs by also acquiring on Philips and Siemens. N.B.: The biggest change is the spatial resolution (2.0 mm isotropic), not the number of gradients. Scanners without multiband capability run the ADNI3 Basic protocol, which is nominally 6 b = 0 + 48 b = 1000. However, the 2.0 mm voxels are taxing for scanners with slower gradient coils, and it was necessary to reduce their number of directions because the scans would have been extremely long otherwise. They have at least 30 b = 1000, though, based on the principle that although more directions are always better, for b = 1000 the improvement in quality slows down after 30. (Derek Jones et al, 2009, IIRC) Essentially each direction gets convolved with its angular response function, which is fat enough at b = 1000 for 30 to provide good coverage. Thus the difference between 41 and 48 b = 1000 is minimal unless you hit one of the magic spherical harmonic numbers, e.g. 45. The highest supported SH order is usually very noisy anyway, though. The *real* difference between these scans is going to be from the receiver coils and TE. TR is so long that the spins are fully relaxed, so it can be ignored. Scanners with multiband capability run the ADNI3 Advanced protocol, which has b = 500, 1000, and 2000 shells. (Prismas only at the start of ADNI3, but this gradually grew to include Philips and other Siemens models, and will include GE in ADNI4). Note that the b = 1000 shell is the same as in the ADNI3 Basic protocol. Whether you want to strictly filter in order to approach homogeneity, or rely on harmonization, or (my personal recommendation) something in between, is up to you. The idea of ADNI selecting "curated" collections for user convenience has been discussed, but that would be most applicable to ADNI3, which is still ongoing. I hope this helps, Rob
Hello Surendra, An alternative approach to RESTORE is to detect outliers on a slicewise basis and either exclude them from the fit (back to RESTORE, basically), or replace them with interpolated values (FSL's eddy_cuda), which also works with DKI, NODDI, etc.. Detecting outliers slicewise tends to be cleaner/more reliable since that's how they happen - motion or whatever corrupts a slice=> the whole slice is corrupted. I'm also hinting at using eddy because I noticed you're correcting for head motion and eddy currents with affine registration. Usually affine registrations of b>0 to b=0 will produce a magnification bias, depending on the cost function. That can be corrected using a model that accounts for both the diffusion contrast and the fact that isotropic magnification is not wanted. Best wishes, Rob On Wed, Nov 10, 2021, 2:34 PM Maharjan, Surendra <smaharj@iu.edu> wrote:
Hello Dipy Community, I have been running DKI fitting using ‘restore’ method. However, I get this message. I am waiting for your favorable response as soon as possible. It says something about the line 1891 in dti.py. https://github.com/dipy/dipy/blob/master/dipy/reconst/dti.py Many Thanks, Suren [Text Description automatically generated] From: Maharjan, Surendra <smaharj@iu.edu> Date: Wednesday, November 10, 2021 at 1:23 PM To: dipy@python.org <dipy@python.org> Subject: DTI and DKI processing Hello Dipy Developers, My name is Suren. I am doing group-wise analysis of Alzheimer’s disease and healthy controls normal brain. I could successfully run DTI and DKI analysis. I did all other preprocessing steps: 1. Denoising (dwidenoise from MRtrix3) 2. Gibb’s artifact removal (from MRtrix3) 3. Eddy current correction and motion artifact removal (affine registration) Yet, I am wondering if we need data normalization and data regularization before performing DTI and DKI analysis. Also, I am wondering if we need to remove outliers before conducting both. I am using 3 shell dMRI data acquired in 9.4 Tesla MRI Scanner. I am waiting for your favorable response. Thank you very much. Best Regards, Suren Indiana University
Hi Suren, These warnings usually happen when your mask happens to include some voxels that are not in the brain (e.g., from the skull or the background surrounding the brain). Nothing to worry about, in almost all cases. Cheers, Ariel On Wed, Nov 24, 2021 at 5:36 PM Maharjan, Surendra <smaharj@iu.edu> wrote:
Hi Ariel, Many Thanks for your valuable response. I am wondering why there’s null areas in the corpus callosum in the mice brain here in this image. This is the DKI- Mean Kurtosis Map. I used ‘weighted least square’, ‘ordinary least square’, ‘restore’, ‘non-linear least square’. All the fit methods have similar null signal areas in the corpus callosum. However, Mean Diffusivity (MD) is fine. The b-values from 0 to 7000, multi-shell data with 0.1 voxel size, multi-shot EPI sequence. [cid:image001.png@01D7E182.DB7A6A70] MK [A picture containing clothing Description automatically generated] MD Many Thanks, Suren From: Ariel Rokem <arokem@uw.edu> Date: Wednesday, November 24, 2021 at 10:08 PM To: Maharjan, Surendra <smaharj@iu.edu> Cc: dipy@python.org <dipy@python.org> Subject: [External] Re: [DIPY] Re: DTI and DKI processing This message was sent from a non-IU address. Please exercise caution when clicking links or opening attachments from external sources. Hi Suren, These warnings usually happen when your mask happens to include some voxels that are not in the brain (e.g., from the skull or the background surrounding the brain). Nothing to worry about, in almost all cases. Cheers, Ariel On Wed, Nov 24, 2021 at 5:36 PM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hello Dipy Community, I have been running DKI fitting using ‘restore’ method. However, I get this message. I am waiting for your favorable response as soon as possible. It says something about the line 1891 in dti.py. https://github.com/dipy/dipy/blob/master/dipy/reconst/dti.py Many Thanks, Suren [Text Description automatically generated] From: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Date: Wednesday, November 10, 2021 at 1:23 PM To: dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: DTI and DKI processing Hello Dipy Developers, My name is Suren. I am doing group-wise analysis of Alzheimer’s disease and healthy controls normal brain. I could successfully run DTI and DKI analysis. I did all other preprocessing steps: 1. Denoising (dwidenoise from MRtrix3) 2. Gibb’s artifact removal (from MRtrix3) 3. Eddy current correction and motion artifact removal (affine registration) Yet, I am wondering if we need data normalization and data regularization before performing DTI and DKI analysis. Also, I am wondering if we need to remove outliers before conducting both. I am using 3 shell dMRI data acquired in 9.4 Tesla MRI Scanner. I am waiting for your favorable response. Thank you very much. Best Regards, Suren Indiana University _______________________________________________ DIPY mailing list -- dipy@python.org<mailto:dipy@python.org> To unsubscribe send an email to dipy-leave@python.org<mailto:dipy-leave@python.org> https://mail.python.org/mailman3/lists/dipy.python.org/ Member address: arokem@gmail.com<mailto:arokem@gmail.com>
Hi Suren, These are usually issues related to noise in the data. Remind me again: What are you doing to denoise this data? We saw similar things with some settings of Patch2Self, but that was resolved in https://github.com/dipy/dipy/issues/2334. Cheers, Ariel On Wed, Nov 24, 2021 at 7:31 PM Maharjan, Surendra <smaharj@iu.edu> wrote:
If you aren't already, I'd recommend using Patch2Self, but make sure you are using the settings described in that issue (or a recent version of DIPY, that includes the fixes described therein). Ariel On Wed, Nov 24, 2021 at 7:51 PM Ariel Rokem <arokem@uw.edu> wrote:
Hello Ariel, Yes, I am using the default settings of patch2self. May I know the optimal value for alpha? Many Thanks, Suren From: Ariel Rokem <arokem@uw.edu> Date: Wednesday, November 24, 2021 at 10:55 PM To: Maharjan, Surendra <smaharj@iu.edu> Cc: dipy@python.org <dipy@python.org> Subject: Re: [External] Re: [DIPY] Re: DTI and DKI processing If you aren't already, I'd recommend using Patch2Self, but make sure you are using the settings described in that issue (or a recent version of DIPY, that includes the fixes described therein). Ariel On Wed, Nov 24, 2021 at 7:51 PM Ariel Rokem <arokem@uw.edu<mailto:arokem@uw.edu>> wrote: Hi Suren, These are usually issues related to noise in the data. Remind me again: What are you doing to denoise this data? We saw similar things with some settings of Patch2Self, but that was resolved in https://github.com/dipy/dipy/issues/2334. Cheers, Ariel On Wed, Nov 24, 2021 at 7:31 PM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hi Ariel, Many Thanks for your valuable response. I am wondering why there’s null areas in the corpus callosum in the mice brain here in this image. This is the DKI- Mean Kurtosis Map. I used ‘weighted least square’, ‘ordinary least square’, ‘restore’, ‘non-linear least square’. All the fit methods have similar null signal areas in the corpus callosum. However, Mean Diffusivity (MD) is fine. The b-values from 0 to 7000, multi-shell data with 0.1 voxel size, multi-shot EPI sequence. [cid:image001.png@01D7E188.E921D770] MK [A picture containing clothing Description automatically generated] MD Many Thanks, Suren From: Ariel Rokem <arokem@uw.edu<mailto:arokem@uw.edu>> Date: Wednesday, November 24, 2021 at 10:08 PM To: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Cc: dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: [External] Re: [DIPY] Re: DTI and DKI processing This message was sent from a non-IU address. Please exercise caution when clicking links or opening attachments from external sources. Hi Suren, These warnings usually happen when your mask happens to include some voxels that are not in the brain (e.g., from the skull or the background surrounding the brain). Nothing to worry about, in almost all cases. Cheers, Ariel On Wed, Nov 24, 2021 at 5:36 PM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hello Dipy Community, I have been running DKI fitting using ‘restore’ method. However, I get this message. I am waiting for your favorable response as soon as possible. It says something about the line 1891 in dti.py. https://github.com/dipy/dipy/blob/master/dipy/reconst/dti.py Many Thanks, Suren [Text Description automatically generated] From: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Date: Wednesday, November 10, 2021 at 1:23 PM To: dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: DTI and DKI processing Hello Dipy Developers, My name is Suren. I am doing group-wise analysis of Alzheimer’s disease and healthy controls normal brain. I could successfully run DTI and DKI analysis. I did all other preprocessing steps: 1. Denoising (dwidenoise from MRtrix3) 2. Gibb’s artifact removal (from MRtrix3) 3. Eddy current correction and motion artifact removal (affine registration) Yet, I am wondering if we need data normalization and data regularization before performing DTI and DKI analysis. Also, I am wondering if we need to remove outliers before conducting both. I am using 3 shell dMRI data acquired in 9.4 Tesla MRI Scanner. I am waiting for your favorable response. Thank you very much. Best Regards, Suren Indiana University _______________________________________________ DIPY mailing list -- dipy@python.org<mailto:dipy@python.org> To unsubscribe send an email to dipy-leave@python.org<mailto:dipy-leave@python.org> https://mail.python.org/mailman3/lists/dipy.python.org/ Member address: arokem@gmail.com<mailto:arokem@gmail.com>
Hi Surendra, I don't think you should need to tune the alpha parameter in Patch2Self. Make sure that you are using DIPY 1.4.1. In DIPY >= 1.4.1, the default model is OLS, which will not require using alpha. If you still run into this issue, I am happy to help. I have worked on this issue in the past :) Thanks, Shreyas On Wed, Nov 24, 2021 at 11:18 PM Maharjan, Surendra <smaharj@iu.edu> wrote:
Hello Shreyas, I used like this: denoise = patch2self(data, bvals, model='ols', shift_intensity=True, clip_negative_vals=False, b0_threshold=100, b0_denoising=False) I tried b0_denoising both True and False, still gives me the black areas in the corpus callosum. Fitting is as follows: dki_model = DiffusionKurtosisModel(gtab, fit_method='WLS') dki_fit = dki_model.fit(data, mask=mask) dki_mk = dki_fit.mk() Many Thanks, Suren From: Shreyas Fadnavis <shreyasfadnavis@gmail.com> Date: Thursday, November 25, 2021 at 12:17 AM To: Maharjan, Surendra <smaharj@iu.edu> Cc: Ariel Rokem <arokem@uw.edu>, dipy@python.org <dipy@python.org> Subject: Re: [DIPY] Re: [External] Re: Re: DTI and DKI processing Hi Surendra, I don't think you should need to tune the alpha parameter in Patch2Self. Make sure that you are using DIPY 1.4.1. In DIPY >= 1.4.1, the default model is OLS, which will not require using alpha. If you still run into this issue, I am happy to help. I have worked on this issue in the past :) Thanks, Shreyas On Wed, Nov 24, 2021 at 11:18 PM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hello Ariel, Yes, I am using the default settings of patch2self. May I know the optimal value for alpha? Many Thanks, Suren From: Ariel Rokem <arokem@uw.edu<mailto:arokem@uw.edu>> Date: Wednesday, November 24, 2021 at 10:55 PM To: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Cc: dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: Re: [External] Re: [DIPY] Re: DTI and DKI processing If you aren't already, I'd recommend using Patch2Self, but make sure you are using the settings described in that issue (or a recent version of DIPY, that includes the fixes described therein). Ariel On Wed, Nov 24, 2021 at 7:51 PM Ariel Rokem <arokem@uw.edu<mailto:arokem@uw.edu>> wrote: Hi Suren, These are usually issues related to noise in the data. Remind me again: What are you doing to denoise this data? We saw similar things with some settings of Patch2Self, but that was resolved in https://github.com/dipy/dipy/issues/2334. Cheers, Ariel On Wed, Nov 24, 2021 at 7:31 PM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hi Ariel, Many Thanks for your valuable response. I am wondering why there’s null areas in the corpus callosum in the mice brain here in this image. This is the DKI- Mean Kurtosis Map. I used ‘weighted least square’, ‘ordinary least square’, ‘restore’, ‘non-linear least square’. All the fit methods have similar null signal areas in the corpus callosum. However, Mean Diffusivity (MD) is fine. The b-values from 0 to 7000, multi-shell data with 0.1 voxel size, multi-shot EPI sequence. [cid:image001.png@01D7E192.990140F0] MK [A picture containing clothing Description automatically generated] MD Many Thanks, Suren From: Ariel Rokem <arokem@uw.edu<mailto:arokem@uw.edu>> Date: Wednesday, November 24, 2021 at 10:08 PM To: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Cc: dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: [External] Re: [DIPY] Re: DTI and DKI processing This message was sent from a non-IU address. Please exercise caution when clicking links or opening attachments from external sources. Hi Suren, These warnings usually happen when your mask happens to include some voxels that are not in the brain (e.g., from the skull or the background surrounding the brain). Nothing to worry about, in almost all cases. Cheers, Ariel On Wed, Nov 24, 2021 at 5:36 PM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hello Dipy Community, I have been running DKI fitting using ‘restore’ method. However, I get this message. I am waiting for your favorable response as soon as possible. It says something about the line 1891 in dti.py. https://github.com/dipy/dipy/blob/master/dipy/reconst/dti.py Many Thanks, Suren [Text Description automatically generated] From: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Date: Wednesday, November 10, 2021 at 1:23 PM To: dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: DTI and DKI processing Hello Dipy Developers, My name is Suren. I am doing group-wise analysis of Alzheimer’s disease and healthy controls normal brain. I could successfully run DTI and DKI analysis. I did all other preprocessing steps: 1. Denoising (dwidenoise from MRtrix3) 2. Gibb’s artifact removal (from MRtrix3) 3. Eddy current correction and motion artifact removal (affine registration) Yet, I am wondering if we need data normalization and data regularization before performing DTI and DKI analysis. Also, I am wondering if we need to remove outliers before conducting both. I am using 3 shell dMRI data acquired in 9.4 Tesla MRI Scanner. I am waiting for your favorable response. Thank you very much. Best Regards, Suren Indiana University _______________________________________________ DIPY mailing list -- dipy@python.org<mailto:dipy@python.org> To unsubscribe send an email to dipy-leave@python.org<mailto:dipy-leave@python.org> https://mail.python.org/mailman3/lists/dipy.python.org/ Member address: arokem@gmail.com<mailto:arokem@gmail.com> _______________________________________________ DIPY mailing list -- dipy@python.org<mailto:dipy@python.org> To unsubscribe send an email to dipy-leave@python.org<mailto:dipy-leave@python.org> https://mail.python.org/mailman3/lists/dipy.python.org/ Member address: shreyasfadnavis@gmail.com<mailto:shreyasfadnavis@gmail.com>
Hi Surendra, My guess is that this may be an issue with DKI modelling itself because DKI is not supposed to be used on data that has bvals >= 3000. I still want to try it out on my end. Would it be possible to share NIFTI, BVEC and BVAL files with me (only -- will not share publicly)? I can test it out for you! Thanks, Shreyas On Thu, Nov 25, 2021 at 12:23 AM Maharjan, Surendra <smaharj@iu.edu> wrote:
Hi Surendra, I don't think the degeneracy in the fitting of DKI is stemming from Patch2Self but is from the DKI model itself. I suggest a possible workaround -- Use Patch2Self + Mean Signal DKI The Mean Signal Kurtosis can be obtained with no degeneracies as seen below: [cid:5894dc89-34e8-455d-a9d7-3c7fcd160be2] The tutorial for MSDKI can be found here: https://dipy.org/documentation/1.4.1./examples_built/reconst_msdki/#example-... The relevant math and details can be found in the following papers: 1. https://onlinelibrary.wiley.com/doi/pdf/10.1002/mrm.27606 2. https://www.frontiersin.org/articles/10.3389/fnhum.2021.675433/full Technically, MSK and MK should give you the same information ?? Thanks, Shreyas PS: Just so you know, DKI needs bvals < 3000. As per Ariel's suggestion, do try out using only those gradient directions with bvals < 3000. ________________________________ From: Ariel Rokem <arokem@uw.edu> Sent: Thursday, November 25, 2021 12:51 AM To: Shreyas Fadnavis <shreyasfadnavis@gmail.com> Cc: Maharjan, Surendra <smaharj@iu.edu>; dipy@python.org <dipy@python.org> Subject: [DIPY] Re: [External] Re: Re: DTI and DKI processing On Wed, Nov 24, 2021 at 9:31 PM Shreyas Fadnavis <shreyasfadnavis@gmail.com<mailto:shreyasfadnavis@gmail.com>> wrote: Hi Surendra, My guess is that this may be an issue with DKI modelling itself because DKI is not supposed to be used on data that has bvals >= 3000. That's a really good point! You might want to try this restricting to only b <= 3000, just to see whether that solves the problem. Ariel I still want to try it out on my end. Would it be possible to share NIFTI, BVEC and BVAL files with me (only -- will not share publicly)? I can test it out for you! Thanks, Shreyas On Thu, Nov 25, 2021 at 12:23 AM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hello Shreyas, I used like this: denoise = patch2self(data, bvals, model='ols', shift_intensity=True, clip_negative_vals=False, b0_threshold=100, b0_denoising=False) I tried b0_denoising both True and False, still gives me the black areas in the corpus callosum. Fitting is as follows: dki_model = DiffusionKurtosisModel(gtab, fit_method='WLS') dki_fit = dki_model.fit(data, mask=mask) dki_mk = dki_fit.mk<http://dki_fit.mk>() Many Thanks, Suren From: Shreyas Fadnavis <shreyasfadnavis@gmail.com<mailto:shreyasfadnavis@gmail.com>> Date: Thursday, November 25, 2021 at 12:17 AM To: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Cc: Ariel Rokem <arokem@uw.edu<mailto:arokem@uw.edu>>, dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: Re: [DIPY] Re: [External] Re: Re: DTI and DKI processing Hi Surendra, I don't think you should need to tune the alpha parameter in Patch2Self. Make sure that you are using DIPY 1.4.1. In DIPY >= 1.4.1, the default model is OLS, which will not require using alpha. If you still run into this issue, I am happy to help. I have worked on this issue in the past :) Thanks, Shreyas On Wed, Nov 24, 2021 at 11:18 PM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hello Ariel, Yes, I am using the default settings of patch2self. May I know the optimal value for alpha? Many Thanks, Suren From: Ariel Rokem <arokem@uw.edu<mailto:arokem@uw.edu>> Date: Wednesday, November 24, 2021 at 10:55 PM To: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Cc: dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: Re: [External] Re: [DIPY] Re: DTI and DKI processing If you aren't already, I'd recommend using Patch2Self, but make sure you are using the settings described in that issue (or a recent version of DIPY, that includes the fixes described therein). Ariel On Wed, Nov 24, 2021 at 7:51 PM Ariel Rokem <arokem@uw.edu<mailto:arokem@uw.edu>> wrote: Hi Suren, These are usually issues related to noise in the data. Remind me again: What are you doing to denoise this data? We saw similar things with some settings of Patch2Self, but that was resolved in https://github.com/dipy/dipy/issues/2334. Cheers, Ariel On Wed, Nov 24, 2021 at 7:31 PM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hi Ariel, Many Thanks for your valuable response. I am wondering why there’s null areas in the corpus callosum in the mice brain here in this image. This is the DKI- Mean Kurtosis Map. I used ‘weighted least square’, ‘ordinary least square’, ‘restore’, ‘non-linear least square’. All the fit methods have similar null signal areas in the corpus callosum. However, Mean Diffusivity (MD) is fine. The b-values from 0 to 7000, multi-shell data with 0.1 voxel size, multi-shot EPI sequence. [cid:17d558fb0554cff311] MK [A picture containing clothing Description automatically generated] MD Many Thanks, Suren From: Ariel Rokem <arokem@uw.edu<mailto:arokem@uw.edu>> Date: Wednesday, November 24, 2021 at 10:08 PM To: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Cc: dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: [External] Re: [DIPY] Re: DTI and DKI processing This message was sent from a non-IU address. Please exercise caution when clicking links or opening attachments from external sources. Hi Suren, These warnings usually happen when your mask happens to include some voxels that are not in the brain (e.g., from the skull or the background surrounding the brain). Nothing to worry about, in almost all cases. Cheers, Ariel On Wed, Nov 24, 2021 at 5:36 PM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hello Dipy Community, I have been running DKI fitting using ‘restore’ method. However, I get this message. I am waiting for your favorable response as soon as possible. It says something about the line 1891 in dti.py. https://github.com/dipy/dipy/blob/master/dipy/reconst/dti.py Many Thanks, Suren [Text Description automatically generated] From: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Date: Wednesday, November 10, 2021 at 1:23 PM To: dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: DTI and DKI processing Hello Dipy Developers, My name is Suren. I am doing group-wise analysis of Alzheimer’s disease and healthy controls normal brain. I could successfully run DTI and DKI analysis. I did all other preprocessing steps: 1. Denoising (dwidenoise from MRtrix3) 2. Gibb’s artifact removal (from MRtrix3) 3. Eddy current correction and motion artifact removal (affine registration) Yet, I am wondering if we need data normalization and data regularization before performing DTI and DKI analysis. Also, I am wondering if we need to remove outliers before conducting both. I am using 3 shell dMRI data acquired in 9.4 Tesla MRI Scanner. I am waiting for your favorable response. Thank you very much. Best Regards, Suren Indiana University _______________________________________________ DIPY mailing list -- dipy@python.org<mailto:dipy@python.org> To unsubscribe send an email to dipy-leave@python.org<mailto:dipy-leave@python.org> https://mail.python.org/mailman3/lists/dipy.python.org/ Member address: arokem@gmail.com<mailto:arokem@gmail.com> _______________________________________________ DIPY mailing list -- dipy@python.org<mailto:dipy@python.org> To unsubscribe send an email to dipy-leave@python.org<mailto:dipy-leave@python.org> https://mail.python.org/mailman3/lists/dipy.python.org/ Member address: shreyasfadnavis@gmail.com<mailto:shreyasfadnavis@gmail.com>
Hello Shreyas, Many Many Thanksgiving. Yes, the mean signal kurtosis (MSK) image looks nice and there are no null signal areas in CC. Yet, I am wondering how I can calculate MK, AK and RK from MSK (msdki) model. I really appreciate your time and effort. Best, Suren From: Fadnavis, Shreyas Sanjeev <shfadn@iu.edu> Date: Thursday, November 25, 2021 at 6:49 AM To: Ariel Rokem <arokem@uw.edu>, Shreyas Fadnavis <shreyasfadnavis@gmail.com>, Maharjan, Surendra <smaharj@iu.edu> Cc: dipy@python.org <dipy@python.org> Subject: Re: [DIPY] Re: [External] Re: Re: DTI and DKI processing Hi Surendra, I don't think the degeneracy in the fitting of DKI is stemming from Patch2Self but is from the DKI model itself. I suggest a possible workaround -- Use Patch2Self + Mean Signal DKI The Mean Signal Kurtosis can be obtained with no degeneracies as seen below: [cid:image001.png@01D7E1E4.38E19560] The tutorial for MSDKI can be found here: https://dipy.org/documentation/1.4.1./examples_built/reconst_msdki/#example-... The relevant math and details can be found in the following papers: 1. https://onlinelibrary.wiley.com/doi/pdf/10.1002/mrm.27606 2. https://www.frontiersin.org/articles/10.3389/fnhum.2021.675433/full Technically, MSK and MK should give you the same information 🙂 Thanks, Shreyas PS: Just so you know, DKI needs bvals < 3000. As per Ariel's suggestion, do try out using only those gradient directions with bvals < 3000. ________________________________ From: Ariel Rokem <arokem@uw.edu> Sent: Thursday, November 25, 2021 12:51 AM To: Shreyas Fadnavis <shreyasfadnavis@gmail.com> Cc: Maharjan, Surendra <smaharj@iu.edu>; dipy@python.org <dipy@python.org> Subject: [DIPY] Re: [External] Re: Re: DTI and DKI processing On Wed, Nov 24, 2021 at 9:31 PM Shreyas Fadnavis <shreyasfadnavis@gmail.com<mailto:shreyasfadnavis@gmail.com>> wrote: Hi Surendra, My guess is that this may be an issue with DKI modelling itself because DKI is not supposed to be used on data that has bvals >= 3000. That's a really good point! You might want to try this restricting to only b <= 3000, just to see whether that solves the problem. Ariel I still want to try it out on my end. Would it be possible to share NIFTI, BVEC and BVAL files with me (only -- will not share publicly)? I can test it out for you! Thanks, Shreyas On Thu, Nov 25, 2021 at 12:23 AM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hello Shreyas, I used like this: denoise = patch2self(data, bvals, model='ols', shift_intensity=True, clip_negative_vals=False, b0_threshold=100, b0_denoising=False) I tried b0_denoising both True and False, still gives me the black areas in the corpus callosum. Fitting is as follows: dki_model = DiffusionKurtosisModel(gtab, fit_method='WLS') dki_fit = dki_model.fit(data, mask=mask) dki_mk = dki_fit.mk<http://dki_fit.mk>() Many Thanks, Suren From: Shreyas Fadnavis <shreyasfadnavis@gmail.com<mailto:shreyasfadnavis@gmail.com>> Date: Thursday, November 25, 2021 at 12:17 AM To: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Cc: Ariel Rokem <arokem@uw.edu<mailto:arokem@uw.edu>>, dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: Re: [DIPY] Re: [External] Re: Re: DTI and DKI processing Hi Surendra, I don't think you should need to tune the alpha parameter in Patch2Self. Make sure that you are using DIPY 1.4.1. In DIPY >= 1.4.1, the default model is OLS, which will not require using alpha. If you still run into this issue, I am happy to help. I have worked on this issue in the past :) Thanks, Shreyas On Wed, Nov 24, 2021 at 11:18 PM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hello Ariel, Yes, I am using the default settings of patch2self. May I know the optimal value for alpha? Many Thanks, Suren From: Ariel Rokem <arokem@uw.edu<mailto:arokem@uw.edu>> Date: Wednesday, November 24, 2021 at 10:55 PM To: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Cc: dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: Re: [External] Re: [DIPY] Re: DTI and DKI processing If you aren't already, I'd recommend using Patch2Self, but make sure you are using the settings described in that issue (or a recent version of DIPY, that includes the fixes described therein). Ariel On Wed, Nov 24, 2021 at 7:51 PM Ariel Rokem <arokem@uw.edu<mailto:arokem@uw.edu>> wrote: Hi Suren, These are usually issues related to noise in the data. Remind me again: What are you doing to denoise this data? We saw similar things with some settings of Patch2Self, but that was resolved in https://github.com/dipy/dipy/issues/2334. Cheers, Ariel On Wed, Nov 24, 2021 at 7:31 PM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hi Ariel, Many Thanks for your valuable response. I am wondering why there’s null areas in the corpus callosum in the mice brain here in this image. This is the DKI- Mean Kurtosis Map. I used ‘weighted least square’, ‘ordinary least square’, ‘restore’, ‘non-linear least square’. All the fit methods have similar null signal areas in the corpus callosum. However, Mean Diffusivity (MD) is fine. The b-values from 0 to 7000, multi-shell data with 0.1 voxel size, multi-shot EPI sequence. [cid:image002.png@01D7E1E4.38E19560] MK [A picture containing clothing Description automatically generated] MD Many Thanks, Suren From: Ariel Rokem <arokem@uw.edu<mailto:arokem@uw.edu>> Date: Wednesday, November 24, 2021 at 10:08 PM To: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Cc: dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: [External] Re: [DIPY] Re: DTI and DKI processing This message was sent from a non-IU address. Please exercise caution when clicking links or opening attachments from external sources. Hi Suren, These warnings usually happen when your mask happens to include some voxels that are not in the brain (e.g., from the skull or the background surrounding the brain). Nothing to worry about, in almost all cases. Cheers, Ariel On Wed, Nov 24, 2021 at 5:36 PM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hello Dipy Community, I have been running DKI fitting using ‘restore’ method. However, I get this message. I am waiting for your favorable response as soon as possible. It says something about the line 1891 in dti.py. https://github.com/dipy/dipy/blob/master/dipy/reconst/dti.py Many Thanks, Suren [Text Description automatically generated] From: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Date: Wednesday, November 10, 2021 at 1:23 PM To: dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: DTI and DKI processing Hello Dipy Developers, My name is Suren. I am doing group-wise analysis of Alzheimer’s disease and healthy controls normal brain. I could successfully run DTI and DKI analysis. I did all other preprocessing steps: 1. Denoising (dwidenoise from MRtrix3) 2. Gibb’s artifact removal (from MRtrix3) 3. Eddy current correction and motion artifact removal (affine registration) Yet, I am wondering if we need data normalization and data regularization before performing DTI and DKI analysis. Also, I am wondering if we need to remove outliers before conducting both. I am using 3 shell dMRI data acquired in 9.4 Tesla MRI Scanner. I am waiting for your favorable response. Thank you very much. Best Regards, Suren Indiana University _______________________________________________ DIPY mailing list -- dipy@python.org<mailto:dipy@python.org> To unsubscribe send an email to dipy-leave@python.org<mailto:dipy-leave@python.org> https://mail.python.org/mailman3/lists/dipy.python.org/ Member address: arokem@gmail.com<mailto:arokem@gmail.com> _______________________________________________ DIPY mailing list -- dipy@python.org<mailto:dipy@python.org> To unsubscribe send an email to dipy-leave@python.org<mailto:dipy-leave@python.org> https://mail.python.org/mailman3/lists/dipy.python.org/ Member address: shreyasfadnavis@gmail.com<mailto:shreyasfadnavis@gmail.com>
Hi Suren, I find your processed DKI maps quite high quality. Is this an ex-vivo scan? Do you mind sharing your acquired protocol and scan time, the field strength, the coil info? Thank you! Best wishes Xiang On Thu, 25 Nov 2021 at 14:54, Maharjan, Surendra <smaharj@iu.edu> wrote:
Hello Xiang, Yes, these are ex-vivo scans. Field Strength: 9.4 T b-value: attached Voxel size: 0.1 Scan Protocol: Multishot EPI (EPI factor 8) No reversed phase encoding steps. Phase Encoding: L to R Scan Time:will tell you tomorrow Coil: Phase array coil 2x2 (without parallel imaging, SENSE) Many Thanks, Suren From: Xiang Feng <fengxiang.mr@gmail.com> Date: Thursday, November 25, 2021 at 12:25 AM To: Maharjan, Surendra <smaharj@iu.edu> Cc: Ariel Rokem <arokem@uw.edu>, dipy@python.org <dipy@python.org> Subject: Re: [DIPY] Re: [External] Re: Re: DTI and DKI processing Hi Suren, I find your processed DKI maps quite high quality. Is this an ex-vivo scan? Do you mind sharing your acquired protocol and scan time, the field strength, the coil info? Thank you! Best wishes Xiang On Thu, 25 Nov 2021 at 14:54, Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hi Ariel, Many Thanks for your valuable response. I am wondering why there’s null areas in the corpus callosum in the mice brain here in this image. This is the DKI- Mean Kurtosis Map. I used ‘weighted least square’, ‘ordinary least square’, ‘restore’, ‘non-linear least square’. All the fit methods have similar null signal areas in the corpus callosum. However, Mean Diffusivity (MD) is fine. The b-values from 0 to 7000, multi-shell data with 0.1 voxel size, multi-shot EPI sequence. [cid:image001.png@01D7E193.DCDD1F00] MK [A picture containing clothing Description automatically generated] MD Many Thanks, Suren From: Ariel Rokem <arokem@uw.edu<mailto:arokem@uw.edu>> Date: Wednesday, November 24, 2021 at 10:08 PM To: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Cc: dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: [External] Re: [DIPY] Re: DTI and DKI processing This message was sent from a non-IU address. Please exercise caution when clicking links or opening attachments from external sources. Hi Suren, These warnings usually happen when your mask happens to include some voxels that are not in the brain (e.g., from the skull or the background surrounding the brain). Nothing to worry about, in almost all cases. Cheers, Ariel On Wed, Nov 24, 2021 at 5:36 PM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hello Dipy Community, I have been running DKI fitting using ‘restore’ method. However, I get this message. I am waiting for your favorable response as soon as possible. It says something about the line 1891 in dti.py. https://github.com/dipy/dipy/blob/master/dipy/reconst/dti.py Many Thanks, Suren [Text Description automatically generated] From: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Date: Wednesday, November 10, 2021 at 1:23 PM To: dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: DTI and DKI processing Hello Dipy Developers, My name is Suren. I am doing group-wise analysis of Alzheimer’s disease and healthy controls normal brain. I could successfully run DTI and DKI analysis. I did all other preprocessing steps: 1. Denoising (dwidenoise from MRtrix3) 2. Gibb’s artifact removal (from MRtrix3) 3. Eddy current correction and motion artifact removal (affine registration) Yet, I am wondering if we need data normalization and data regularization before performing DTI and DKI analysis. Also, I am wondering if we need to remove outliers before conducting both. I am using 3 shell dMRI data acquired in 9.4 Tesla MRI Scanner. I am waiting for your favorable response. Thank you very much. Best Regards, Suren Indiana University _______________________________________________ DIPY mailing list -- dipy@python.org<mailto:dipy@python.org> To unsubscribe send an email to dipy-leave@python.org<mailto:dipy-leave@python.org> https://mail.python.org/mailman3/lists/dipy.python.org/ Member address: arokem@gmail.com<mailto:arokem@gmail.com> _______________________________________________ DIPY mailing list -- dipy@python.org<mailto:dipy@python.org> To unsubscribe send an email to dipy-leave@python.org<mailto:dipy-leave@python.org> https://mail.python.org/mailman3/lists/dipy.python.org/ Member address: fengxiang.mr@gmail.com<mailto:fengxiang.mr@gmail.com>
Hi Suren, I am not 100% sure what "data normalization and regularization" means in this context. One thing you might consider for dealing with outliers is to use the RESTORE method for DTI, which robustifies the model fit against outlying observations. Cheers, Ariel On Wed, Nov 10, 2021 at 10:26 AM Maharjan, Surendra <smaharj@iu.edu> wrote:
Hi Ariel, The data normalization here I would like to know: Can we normalize the DWI data in the range 0 to 1 before fitting? Thank you very much. I will work on RESTORE DTI fit. Regularization here I mean: L1 (Lasso) and L2 (Ridge) regularization. Do we have similar robust fit like RESTORE for DKI? Thank you, Ariel. Your comments always saved me. Best Regards, Suren From: Ariel Rokem <arokem@uw.edu> Date: Wednesday, November 10, 2021 at 1:34 PM To: Maharjan, Surendra <smaharj@iu.edu> Cc: dipy@python.org <dipy@python.org> Subject: [External] Re: [DIPY] DTI and DKI processing This message was sent from a non-IU address. Please exercise caution when clicking links or opening attachments from external sources. Hi Suren, I am not 100% sure what "data normalization and regularization" means in this context. One thing you might consider for dealing with outliers is to use the RESTORE method for DTI, which robustifies the model fit against outlying observations. Cheers, Ariel On Wed, Nov 10, 2021 at 10:26 AM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hello Dipy Developers, My name is Suren. I am doing group-wise analysis of Alzheimer’s disease and healthy controls normal brain. I could successfully run DTI and DKI analysis. I did all other preprocessing steps: 1. Denoising (dwidenoise from MRtrix3) 2. Gibb’s artifact removal (from MRtrix3) 3. Eddy current correction and motion artifact removal (affine registration) Yet, I am wondering if we need data normalization and data regularization before performing DTI and DKI analysis. Also, I am wondering if we need to remove outliers before conducting both. I am using 3 shell dMRI data acquired in 9.4 Tesla MRI Scanner. I am waiting for your favorable response. Thank you very much. Best Regards, Suren Indiana University _______________________________________________ DIPY mailing list -- dipy@python.org<mailto:dipy@python.org> To unsubscribe send an email to dipy-leave@python.org<mailto:dipy-leave@python.org> https://mail.python.org/mailman3/lists/dipy.python.org/ Member address: arokem@gmail.com<mailto:arokem@gmail.com>
Hi Suren, On Wed, Nov 10, 2021 at 10:57 AM Maharjan, Surendra <smaharj@iu.edu> wrote:
You can, but that kind of normalization is already baked into DTI and DKI fitting, so it won't change your model fits.
I don't think it makes sense to use these for DTI and/or DKI, which have only a few regressors that are (I think) orthogonal by construction.
Do we have similar robust fit like RESTORE for DKI?
Yeah - I think that you can use RESTORE in DKI as well (initialize the model with `fit_method="RESTORE"` and with a `sigma` value). The more thorough solution for that would be to implement Tax et al.'s REKINDLE method (https://onlinelibrary.wiley.com/doi/full/10.1002/mrm.25165), but we don't have that. Cheers, Ariel
Dear Dipy fellas, My student would like to use many data from the ADNI dataset, but we noticed a lot of heterogeneity. Even taking the data from the same type of scanner and same field (we checked both ADNI1 ADNI2 and ADNI3), there is1. Different TR and TE 2. Different number of gradients We found the COMBAT tool https://github.com/Jfortin1/ComBatHarmonization, which according to my students is a nightmare and also does not solve the issue of different number of gradients. How can we address this?I know a solution would be to drop the ADNI dataset and use something else, but there are other data in this dataset which are not available in other datasets. I know Bramsh Chandio worked on the PPMI dataset which I imagine has the same level of messiness... or not? Best,Alex Prof. Dr. Alessandro Crimi Research Group Leader https://bam.sano.science Centre for Computational Medicine Czarnowiejska 36, building C5. 30-072 Kraków, Poland Phone: +48 575 453 005 www.sano.science ----------------------------------------------- Visiting Lecturer African Institute for Mathematical Sciences www.aims.edu.gh --------------------------------------------------------- https://twitter.com/Dr_Alex_Crimi
Hi Alex, On Thu, Nov 11, 2021 at 5:31 AM Alex Crimi via DIPY <dipy@python.org> wrote:
Dear Dipy fellas,
Hi Alex, Thanks for raising this! It's a tough issue.
I'll let Bramsh chime in about PPMT, but I will mention that we've used COMBAT for harmonization at the level of the diffusion features / tract profiles. We've found it to be rather effective in harmonizing across sites in the HBN study, despite some frustrating differences across sites, though not as much variability in the acquisition protocols (see supplemental S12 here: https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.10091...). This might also address different acquisition parameters to some degree, assuming you have enough subjects in each acquisition protocol. We've found this package to be slightly more user-friendly than the implementation you pointed out: https://github.com/Warvito/neurocombat_sklearn, because it follows the sklearn API. Cheers, Ariel
Hi,thanks for the links. So you confirm you have been able to harmonize given different number of gradient directions?I cannot find this information in your paper.Best,Alex On Thursday, November 11, 2021, 06:34:16 PM GMT+1, Ariel Rokem <arokem@uw.edu> wrote: Hi Alex, On Thu, Nov 11, 2021 at 5:31 AM Alex Crimi via DIPY <dipy@python.org> wrote: Dear Dipy fellas, Hi Alex, Thanks for raising this! It's a tough issue. My student would like to use many data from the ADNI dataset, but we noticed a lot of heterogeneity. Even taking the data from the same type of scanner and same field (we checked both ADNI1 ADNI2 and ADNI3), there is1. Different TR and TE 2. Different number of gradients We found the COMBAT tool https://github.com/Jfortin1/ComBatHarmonization, which according to my students is a nightmare and also does not solve the issue of different number of gradients. How can we address this? I'll let Bramsh chime in about PPMT, but I will mention that we've used COMBAT for harmonization at the level of the diffusion features / tract profiles. We've found it to be rather effective in harmonizing across sites in the HBN study, despite some frustrating differences across sites, though not as much variability in the acquisition protocols (see supplemental S12 here: https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.10091...). This might also address different acquisition parameters to some degree, assuming you have enough subjects in each acquisition protocol. We've found this package to be slightly more user-friendly than the implementation you pointed out: https://github.com/Warvito/neurocombat_sklearn, because it follows the sklearn API. Cheers, Ariel I know a solution would be to drop the ADNI dataset and use something else, but there are other data in this dataset which are not available in other datasets. I know Bramsh Chandio worked on the PPMI dataset which I imagine has the same level of messiness... or not? Best,Alex Prof. Dr. Alessandro Crimi Research Group Leader https://bam.sano.science Centre for Computational Medicine Czarnowiejska 36, building C5. 30-072 Kraków, Poland Phone: +48 575 453 005 www.sano.science ----------------------------------------------- Visiting Lecturer African Institute for Mathematical Sciences www.aims.edu.gh --------------------------------------------------------- https://twitter.com/Dr_Alex_Crimi _______________________________________________ DIPY mailing list -- dipy@python.org To unsubscribe send an email to dipy-leave@python.org https://mail.python.org/mailman3/lists/dipy.python.org/ Member address: arokem@gmail.com _______________________________________________ DIPY mailing list -- dipy@python.org To unsubscribe send an email to dipy-leave@python.org https://mail.python.org/mailman3/lists/dipy.python.org/ Member address: alex.crimi@yahoo.com
Hi Alex, On Thu, Nov 11, 2021 at 9:48 AM Alex Crimi <alex.crimi@yahoo.com> wrote:
No - sorry - the acquisitions in HBN have the same number of gradient directions in all of the sites, so much less variable than what you are facing with ADNI. Cheers, Ariel
Hi Alex, Ariel, Just to add another citation, ADNI3 protocol analyses using ROIs+diffusion features/indices and ComBat can be found here: https://internal-journal.frontiersin.org/articles/10.3389/fninf.2019.00002/f... Supplementary figures 1 and 2 are perhaps the most relevant, which seem to show that ComBat is removing differences for linear associations. Another option though would be to harmonize in the local model parameter space, instead of in the original spaces. Hope this helps, Daniel Moyer On Thu, Nov 11, 2021 at 12:34 PM Ariel Rokem <arokem@uw.edu> wrote:
Hi everyone, Check this out: https://github.com/pnlbwh/dMRIharmonization Best, Suheyla Cetin-Karayumak, PhD Assistant Professor, Psychiatry Neuroimaging Laboratory, Brigham and Women's Hospital, Harvard Medical School, Boston. skarayumak@bwh.harvard.edu ________________________________ From: Ariel Rokem <arokem@uw.edu> Sent: Thursday, November 11, 2021 12:14 PM To: Alex Crimi <alex.crimi@yahoo.com> Cc: dipy@python.org <dipy@python.org>; bqchandi@iu.edu <bqchandi@iu.edu> Subject: [DIPY] Re: Hamonizing ADNI for some Dipy stuff External Email - Use Caution Hi Alex, On Thu, Nov 11, 2021 at 5:31 AM Alex Crimi via DIPY <dipy@python.org<mailto:dipy@python.org>> wrote: Dear Dipy fellas, Hi Alex, Thanks for raising this! It's a tough issue. My student would like to use many data from the ADNI dataset, but we noticed a lot of heterogeneity. Even taking the data from the same type of scanner and same field (we checked both ADNI1 ADNI2 and ADNI3), there is 1. Different TR and TE 2. Different number of gradients We found the COMBAT tool https://github.com/Jfortin1/ComBatHarmonization<https://secure-web.cisco.com/1qYv-oUqGK6oTHJSqWwxscxc52VJxb2jOkmS7yFdecQIJNuuCiSoREhwuTKVlNtzhG3hfvoR5UqiopXrRtqK4Ob68mBw6MpSN4ufv44D8ldLMFslolNQiKN8CMdEWxc5khutYaFynJdS-A-1o-ZdNYfx9wUiBK-fpoEJ1u0pzslcEZdOfSaQddDHw4QMjtDizSCR_7Blx0_rm5djrduVisSGqW69fw74DTAkwngTNm_O-SdladRDKFsczre-pB2EZY6U1gEWrPHZ7RjbKNkhCpVt_XwIDt62eT9Y6A1UX_wIVwgvnkehhSJjI7YRZazGU/https%3A%2F%2Fgithub.com%2FJfortin1%2FComBatHarmonization>, which according to my students is a nightmare and also does not solve the issue of different number of gradients. How can we address this? I'll let Bramsh chime in about PPMT, but I will mention that we've used COMBAT for harmonization at the level of the diffusion features / tract profiles. We've found it to be rather effective in harmonizing across sites in the HBN study, despite some frustrating differences across sites, though not as much variability in the acquisition protocols (see supplemental S12 here: https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1009136<https://secure-web.cisco.com/1yOHjaWE-bwF448WN4jS75U6VClF1g3YxyG8taPUdf1PEZhgDgBXzcizpeDg1yVw5D1eTNs3YyrxBAsSjNbAa9VG8o6qgevdkoqitclMRCd70mDq2bk_fjkkzOud2aiKzkQGLAp_Eyhxq58cCMktJOh8ba0YVN-k0dhLeoW4AB5wpIVLky_bW95PolWjC2QZlt72Ioqmp5Pu_PxBG8bbG6lqwZi5ze1apkvDmI8b7wbGYGapgfzivr-SMvyLStrvmrC517WXYWT5oFQS1J8SZ2AxBjdPi09Kv-pl-UXNh-HoW8NX9BZWypqXybkA-PZp8/https%3A%2F%2Fjournals.plos.org%2Fploscompbiol%2Farticle%3Fid%3D10.1371%2Fjournal.pcbi.1009136>). This might also address different acquisition parameters to some degree, assuming you have enough subjects in each acquisition protocol. We've found this package to be slightly more user-friendly than the implementation you pointed out: https://github.com/Warvito/neurocombat_sklearn<https://secure-web.cisco.com/1gL9DmOOTnMNnyUG6j3HgqAJ_n3YvoN6VZWCe8tE1u2_-VEnrI1s9SGW5GraEY0bT_sjIaxQ9Kku8suxXoi7oXO0D_613QwD4Ldov3MHLYpEGJ99tcL9zGh0pbEwJIZNrl7R5NV3P4u3MOTRzF8k-dxkX5l3R0N1WHsG3XqpqjuvYr02jitrFO_j1WI9FCxT0pBxrZ7yGGQipG_jSeVOunSr9PvSFtclU8F4wy4pSKzBYga_zGGdVnAfejK55xxRNQyl_om1Cs73oLzfs2enGhL2RR2-Fzd7PjKrS6K-RDvdlunflqraBg09PTWQd_Q2V/https%3A%2F%2Fgithub.com%2FWarvito%2Fneurocombat_sklearn>, because it follows the sklearn API. Cheers, Ariel I know a solution would be to drop the ADNI dataset and use something else, but there are other data in this dataset which are not available in other datasets. I know Bramsh Chandio worked on the PPMI dataset which I imagine has the same level of messiness... or not? Best, Alex Prof. Dr. Alessandro Crimi Research Group Leader https://bam.sano.science<https://secure-web.cisco.com/1P37o1AQpD282af3eR09BoxrOLJ9c8TRIL8nF3QyjlWf_80xB1hT3dp3ioKZmBx9gjKBq3JHpBLRLIhiCpdlBWLpLNX0N9hQdN80F3Z-7q5WL33BgjOXno6AaPjS9zmmjOV3ERdkpl1iJnf1qTEtjJMtVU0fKFL_7OdGVdBRx1RCcrGDGe4KUNGO844TliQ2JJJfaZIOv8iqtIBAdLt1nvbPj6t61zChSSNGMEH2BMNph0zjtd6EGbQl2nSJbxlw_RLBUdNigSLh6uOtFyo5OTnW3I26NqdywNVMhsmy3GXF6M9tOfNZQNjBO8rcuxRTM/https%3A%2F%2Fbam.sano.science> [X] Centre for Computational Medicine Czarnowiejska 36, building C5. 30-072 Kraków, Poland Phone: +48 575 453 005 www.sano.science<http://secure-web.cisco.com/1eGjBi4ybQ-RwvzMnhpzMGzXIH3QHrd4acViwk311yDFFOWk...> ----------------------------------------------- Visiting Lecturer African Institute for Mathematical Sciences www.aims.edu.gh<http://secure-web.cisco.com/1rMVeXDkP0BFhRBHJTIklU3_iREWl_6EvFl4ifvYOjSwD7k7...> --------------------------------------------------------- https://twitter.com/Dr_Alex_Crimi<https://secure-web.cisco.com/11JPLwI8c7Da7TrruKifMgFOOC4GbBFxJA1QF3zSjJlTQKcarnuC6-tmQMuUlNqBgE2UQ9n8eE0AoXnjaLIzM2LUYptrmEDAiNi4d0P46sv0IEeJX3gWe6OLurfH20C0vwZk7Y_myAkRewafeUnXKBcShaATXuIjq_4hh4jqFrlU0VDMdFAf8wb4Nmkdpst1ytEharGordws3_xleOA-ECc2dOKJYXctkn-u_iavFBCO5XQ7_i4lGMySfbcfZw7zhUFT382Ce7af5aVIt4WBHZtLjFmOz3OyROnRlsqSuHpyqxlJCm42ng5YOhRdPa0MR/https%3A%2F%2Ftwitter.com%2FDr_Alex_Crimi> _______________________________________________ DIPY mailing list -- dipy@python.org<mailto:dipy@python.org> To unsubscribe send an email to dipy-leave@python.org<mailto:dipy-leave@python.org> https://mail.python.org/mailman3/lists/dipy.python.org/<https://secure-web.cisco.com/1SPfzZRIgTfgpftXLS6ppjvH5VKfjmmP7WUkFeHWx5vEvwJOIt0f0IbIVj1BhTnwxXyGV1F3gn-HbgrL1dKC3JM3hsfwnH7lpCAYS-bF4wfLQVBApGk7NMuy3lsHbOQYBodcwJkImTdNNpdUSnoPU82P8OUYw_YI7VyL900cygT0ER7NiU1ZrJRFub11zINpIvZH5zqa4CAm3Z__WpJaW2swTDM2TmnxiZJXIaUpCAIxhWS9mQmpPHK6WuPpt_MH4GjjQuTEW_WNYu1T4OaYzk-05B-EsxGsJGYkxJPsGt5nRiG5g8m4-03cjk7eMG9xm/https%3A%2F%2Fmail.python.org%2Fmailman3%2Flists%2Fdipy.python.org%2F> Member address: arokem@gmail.com<mailto:arokem@gmail.com> The information in this e-mail is intended only for the person to whom it is addressed. If you believe this e-mail was sent to you in error and the e-mail contains patient information, please contact the Mass General Brigham Compliance HelpLine at http://www.massgeneralbrigham.org/complianceline . If the e-mail was sent to you in error but does not contain patient information, please contact the sender and properly dispose of the e-mail. Please note that this e-mail is not secure (encrypted). If you do not wish to continue communication over unencrypted e-mail, please notify the sender of this message immediately. Continuing to send or respond to e-mail after receiving this message means you understand and accept this risk and wish to continue to communicate over unencrypted e-mail.
Hi everyone, Also check the latest work on harmonization of diffusion MRI on ADNI using COMBAT, COMBAT-GAM and COVBAT by Sophia Thomopoulos. It will be presented at SIPAIM 2021 (https://sipaim.org): https://www.medrxiv.org/content/10.1101/2021.10.04.21263994v1.full.pdf Cheers, Julio El jue, 11 nov 2021 a las 10:42, Cetin Karayumak, Suheyla (< SKARAYUMAK@bwh.harvard.edu>) escribió:
Hello, ADNI questions are welcomed at http://adni.loni.usc.edu/support/experts-knowledge-base/ask-experts/ , where questions are forwarded to the appropriate experts according to their category. We don't expect other forums to provide answers for us, although dipy's has done a fine job. Data harmonization is an active area of research, but as noted in Sophia Thomopoulos' preprint ComBat does a good job for atlas region values, and Daniel Moyer's papers would be a good start for the tougher problem of making the images themselves more directly comparable. One approach that I don't think has been mentioned in this thread is to extract a homogeneous subset from the ADNI data. After all, the reason ADNI has heterogeneity is that it spans multiple scanners, and actually, multiple generations of scanners. If we had stuck to a homogeneous set of scanners the resulting data set would have been about the same size as pulling a homogeneous set out of ADNI. In practice when selecting a set it is best to know which differences can be ignored, which can be harmonized reasonably easily, and which should be avoided. ADNI has had 3 stages so far, each lasting about 5 years. ADNI1: Did not have dMRI. I was not involved with ADNI until much later, but my understanding is that any "ADNI1 DTIs" you might find were experimental and a standard protocol had not been settled on yet. ADNI2: Actually was relatively homogeneous, because it was restricted to GE, with 5 b = 0 + 41 b = 1000, 2.7mm resolution. You might find a few 4 + 41s - I would treat them as the same. You might also find a few with 60 b = 1000. Those were *not* the standard ADNI2 protocol, and typically have a matching standard DTI for the same participant and date. i.e. they're not for people wanting a homogeneous set, they're for people *wanting* to try out their harmonization method. (Even though at the time the main motivation was likely testing out an enhanced number of directions.) ADNI3: Deliberately broke with ADNI2, in order to keep up with advances in the field, and to increase the number of dMRIs by also acquiring on Philips and Siemens. N.B.: The biggest change is the spatial resolution (2.0 mm isotropic), not the number of gradients. Scanners without multiband capability run the ADNI3 Basic protocol, which is nominally 6 b = 0 + 48 b = 1000. However, the 2.0 mm voxels are taxing for scanners with slower gradient coils, and it was necessary to reduce their number of directions because the scans would have been extremely long otherwise. They have at least 30 b = 1000, though, based on the principle that although more directions are always better, for b = 1000 the improvement in quality slows down after 30. (Derek Jones et al, 2009, IIRC) Essentially each direction gets convolved with its angular response function, which is fat enough at b = 1000 for 30 to provide good coverage. Thus the difference between 41 and 48 b = 1000 is minimal unless you hit one of the magic spherical harmonic numbers, e.g. 45. The highest supported SH order is usually very noisy anyway, though. The *real* difference between these scans is going to be from the receiver coils and TE. TR is so long that the spins are fully relaxed, so it can be ignored. Scanners with multiband capability run the ADNI3 Advanced protocol, which has b = 500, 1000, and 2000 shells. (Prismas only at the start of ADNI3, but this gradually grew to include Philips and other Siemens models, and will include GE in ADNI4). Note that the b = 1000 shell is the same as in the ADNI3 Basic protocol. Whether you want to strictly filter in order to approach homogeneity, or rely on harmonization, or (my personal recommendation) something in between, is up to you. The idea of ADNI selecting "curated" collections for user convenience has been discussed, but that would be most applicable to ADNI3, which is still ongoing. I hope this helps, Rob
Hello Surendra, An alternative approach to RESTORE is to detect outliers on a slicewise basis and either exclude them from the fit (back to RESTORE, basically), or replace them with interpolated values (FSL's eddy_cuda), which also works with DKI, NODDI, etc.. Detecting outliers slicewise tends to be cleaner/more reliable since that's how they happen - motion or whatever corrupts a slice=> the whole slice is corrupted. I'm also hinting at using eddy because I noticed you're correcting for head motion and eddy currents with affine registration. Usually affine registrations of b>0 to b=0 will produce a magnification bias, depending on the cost function. That can be corrected using a model that accounts for both the diffusion contrast and the fact that isotropic magnification is not wanted. Best wishes, Rob On Wed, Nov 10, 2021, 2:34 PM Maharjan, Surendra <smaharj@iu.edu> wrote:
Hello Dipy Community, I have been running DKI fitting using ‘restore’ method. However, I get this message. I am waiting for your favorable response as soon as possible. It says something about the line 1891 in dti.py. https://github.com/dipy/dipy/blob/master/dipy/reconst/dti.py Many Thanks, Suren [Text Description automatically generated] From: Maharjan, Surendra <smaharj@iu.edu> Date: Wednesday, November 10, 2021 at 1:23 PM To: dipy@python.org <dipy@python.org> Subject: DTI and DKI processing Hello Dipy Developers, My name is Suren. I am doing group-wise analysis of Alzheimer’s disease and healthy controls normal brain. I could successfully run DTI and DKI analysis. I did all other preprocessing steps: 1. Denoising (dwidenoise from MRtrix3) 2. Gibb’s artifact removal (from MRtrix3) 3. Eddy current correction and motion artifact removal (affine registration) Yet, I am wondering if we need data normalization and data regularization before performing DTI and DKI analysis. Also, I am wondering if we need to remove outliers before conducting both. I am using 3 shell dMRI data acquired in 9.4 Tesla MRI Scanner. I am waiting for your favorable response. Thank you very much. Best Regards, Suren Indiana University
Hi Suren, These warnings usually happen when your mask happens to include some voxels that are not in the brain (e.g., from the skull or the background surrounding the brain). Nothing to worry about, in almost all cases. Cheers, Ariel On Wed, Nov 24, 2021 at 5:36 PM Maharjan, Surendra <smaharj@iu.edu> wrote:
Hi Ariel, Many Thanks for your valuable response. I am wondering why there’s null areas in the corpus callosum in the mice brain here in this image. This is the DKI- Mean Kurtosis Map. I used ‘weighted least square’, ‘ordinary least square’, ‘restore’, ‘non-linear least square’. All the fit methods have similar null signal areas in the corpus callosum. However, Mean Diffusivity (MD) is fine. The b-values from 0 to 7000, multi-shell data with 0.1 voxel size, multi-shot EPI sequence. [cid:image001.png@01D7E182.DB7A6A70] MK [A picture containing clothing Description automatically generated] MD Many Thanks, Suren From: Ariel Rokem <arokem@uw.edu> Date: Wednesday, November 24, 2021 at 10:08 PM To: Maharjan, Surendra <smaharj@iu.edu> Cc: dipy@python.org <dipy@python.org> Subject: [External] Re: [DIPY] Re: DTI and DKI processing This message was sent from a non-IU address. Please exercise caution when clicking links or opening attachments from external sources. Hi Suren, These warnings usually happen when your mask happens to include some voxels that are not in the brain (e.g., from the skull or the background surrounding the brain). Nothing to worry about, in almost all cases. Cheers, Ariel On Wed, Nov 24, 2021 at 5:36 PM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hello Dipy Community, I have been running DKI fitting using ‘restore’ method. However, I get this message. I am waiting for your favorable response as soon as possible. It says something about the line 1891 in dti.py. https://github.com/dipy/dipy/blob/master/dipy/reconst/dti.py Many Thanks, Suren [Text Description automatically generated] From: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Date: Wednesday, November 10, 2021 at 1:23 PM To: dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: DTI and DKI processing Hello Dipy Developers, My name is Suren. I am doing group-wise analysis of Alzheimer’s disease and healthy controls normal brain. I could successfully run DTI and DKI analysis. I did all other preprocessing steps: 1. Denoising (dwidenoise from MRtrix3) 2. Gibb’s artifact removal (from MRtrix3) 3. Eddy current correction and motion artifact removal (affine registration) Yet, I am wondering if we need data normalization and data regularization before performing DTI and DKI analysis. Also, I am wondering if we need to remove outliers before conducting both. I am using 3 shell dMRI data acquired in 9.4 Tesla MRI Scanner. I am waiting for your favorable response. Thank you very much. Best Regards, Suren Indiana University _______________________________________________ DIPY mailing list -- dipy@python.org<mailto:dipy@python.org> To unsubscribe send an email to dipy-leave@python.org<mailto:dipy-leave@python.org> https://mail.python.org/mailman3/lists/dipy.python.org/ Member address: arokem@gmail.com<mailto:arokem@gmail.com>
Hi Suren, These are usually issues related to noise in the data. Remind me again: What are you doing to denoise this data? We saw similar things with some settings of Patch2Self, but that was resolved in https://github.com/dipy/dipy/issues/2334. Cheers, Ariel On Wed, Nov 24, 2021 at 7:31 PM Maharjan, Surendra <smaharj@iu.edu> wrote:
If you aren't already, I'd recommend using Patch2Self, but make sure you are using the settings described in that issue (or a recent version of DIPY, that includes the fixes described therein). Ariel On Wed, Nov 24, 2021 at 7:51 PM Ariel Rokem <arokem@uw.edu> wrote:
Hello Ariel, Yes, I am using the default settings of patch2self. May I know the optimal value for alpha? Many Thanks, Suren From: Ariel Rokem <arokem@uw.edu> Date: Wednesday, November 24, 2021 at 10:55 PM To: Maharjan, Surendra <smaharj@iu.edu> Cc: dipy@python.org <dipy@python.org> Subject: Re: [External] Re: [DIPY] Re: DTI and DKI processing If you aren't already, I'd recommend using Patch2Self, but make sure you are using the settings described in that issue (or a recent version of DIPY, that includes the fixes described therein). Ariel On Wed, Nov 24, 2021 at 7:51 PM Ariel Rokem <arokem@uw.edu<mailto:arokem@uw.edu>> wrote: Hi Suren, These are usually issues related to noise in the data. Remind me again: What are you doing to denoise this data? We saw similar things with some settings of Patch2Self, but that was resolved in https://github.com/dipy/dipy/issues/2334. Cheers, Ariel On Wed, Nov 24, 2021 at 7:31 PM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hi Ariel, Many Thanks for your valuable response. I am wondering why there’s null areas in the corpus callosum in the mice brain here in this image. This is the DKI- Mean Kurtosis Map. I used ‘weighted least square’, ‘ordinary least square’, ‘restore’, ‘non-linear least square’. All the fit methods have similar null signal areas in the corpus callosum. However, Mean Diffusivity (MD) is fine. The b-values from 0 to 7000, multi-shell data with 0.1 voxel size, multi-shot EPI sequence. [cid:image001.png@01D7E188.E921D770] MK [A picture containing clothing Description automatically generated] MD Many Thanks, Suren From: Ariel Rokem <arokem@uw.edu<mailto:arokem@uw.edu>> Date: Wednesday, November 24, 2021 at 10:08 PM To: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Cc: dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: [External] Re: [DIPY] Re: DTI and DKI processing This message was sent from a non-IU address. Please exercise caution when clicking links or opening attachments from external sources. Hi Suren, These warnings usually happen when your mask happens to include some voxels that are not in the brain (e.g., from the skull or the background surrounding the brain). Nothing to worry about, in almost all cases. Cheers, Ariel On Wed, Nov 24, 2021 at 5:36 PM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hello Dipy Community, I have been running DKI fitting using ‘restore’ method. However, I get this message. I am waiting for your favorable response as soon as possible. It says something about the line 1891 in dti.py. https://github.com/dipy/dipy/blob/master/dipy/reconst/dti.py Many Thanks, Suren [Text Description automatically generated] From: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Date: Wednesday, November 10, 2021 at 1:23 PM To: dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: DTI and DKI processing Hello Dipy Developers, My name is Suren. I am doing group-wise analysis of Alzheimer’s disease and healthy controls normal brain. I could successfully run DTI and DKI analysis. I did all other preprocessing steps: 1. Denoising (dwidenoise from MRtrix3) 2. Gibb’s artifact removal (from MRtrix3) 3. Eddy current correction and motion artifact removal (affine registration) Yet, I am wondering if we need data normalization and data regularization before performing DTI and DKI analysis. Also, I am wondering if we need to remove outliers before conducting both. I am using 3 shell dMRI data acquired in 9.4 Tesla MRI Scanner. I am waiting for your favorable response. Thank you very much. Best Regards, Suren Indiana University _______________________________________________ DIPY mailing list -- dipy@python.org<mailto:dipy@python.org> To unsubscribe send an email to dipy-leave@python.org<mailto:dipy-leave@python.org> https://mail.python.org/mailman3/lists/dipy.python.org/ Member address: arokem@gmail.com<mailto:arokem@gmail.com>
Hi Surendra, I don't think you should need to tune the alpha parameter in Patch2Self. Make sure that you are using DIPY 1.4.1. In DIPY >= 1.4.1, the default model is OLS, which will not require using alpha. If you still run into this issue, I am happy to help. I have worked on this issue in the past :) Thanks, Shreyas On Wed, Nov 24, 2021 at 11:18 PM Maharjan, Surendra <smaharj@iu.edu> wrote:
Hello Shreyas, I used like this: denoise = patch2self(data, bvals, model='ols', shift_intensity=True, clip_negative_vals=False, b0_threshold=100, b0_denoising=False) I tried b0_denoising both True and False, still gives me the black areas in the corpus callosum. Fitting is as follows: dki_model = DiffusionKurtosisModel(gtab, fit_method='WLS') dki_fit = dki_model.fit(data, mask=mask) dki_mk = dki_fit.mk() Many Thanks, Suren From: Shreyas Fadnavis <shreyasfadnavis@gmail.com> Date: Thursday, November 25, 2021 at 12:17 AM To: Maharjan, Surendra <smaharj@iu.edu> Cc: Ariel Rokem <arokem@uw.edu>, dipy@python.org <dipy@python.org> Subject: Re: [DIPY] Re: [External] Re: Re: DTI and DKI processing Hi Surendra, I don't think you should need to tune the alpha parameter in Patch2Self. Make sure that you are using DIPY 1.4.1. In DIPY >= 1.4.1, the default model is OLS, which will not require using alpha. If you still run into this issue, I am happy to help. I have worked on this issue in the past :) Thanks, Shreyas On Wed, Nov 24, 2021 at 11:18 PM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hello Ariel, Yes, I am using the default settings of patch2self. May I know the optimal value for alpha? Many Thanks, Suren From: Ariel Rokem <arokem@uw.edu<mailto:arokem@uw.edu>> Date: Wednesday, November 24, 2021 at 10:55 PM To: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Cc: dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: Re: [External] Re: [DIPY] Re: DTI and DKI processing If you aren't already, I'd recommend using Patch2Self, but make sure you are using the settings described in that issue (or a recent version of DIPY, that includes the fixes described therein). Ariel On Wed, Nov 24, 2021 at 7:51 PM Ariel Rokem <arokem@uw.edu<mailto:arokem@uw.edu>> wrote: Hi Suren, These are usually issues related to noise in the data. Remind me again: What are you doing to denoise this data? We saw similar things with some settings of Patch2Self, but that was resolved in https://github.com/dipy/dipy/issues/2334. Cheers, Ariel On Wed, Nov 24, 2021 at 7:31 PM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hi Ariel, Many Thanks for your valuable response. I am wondering why there’s null areas in the corpus callosum in the mice brain here in this image. This is the DKI- Mean Kurtosis Map. I used ‘weighted least square’, ‘ordinary least square’, ‘restore’, ‘non-linear least square’. All the fit methods have similar null signal areas in the corpus callosum. However, Mean Diffusivity (MD) is fine. The b-values from 0 to 7000, multi-shell data with 0.1 voxel size, multi-shot EPI sequence. [cid:image001.png@01D7E192.990140F0] MK [A picture containing clothing Description automatically generated] MD Many Thanks, Suren From: Ariel Rokem <arokem@uw.edu<mailto:arokem@uw.edu>> Date: Wednesday, November 24, 2021 at 10:08 PM To: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Cc: dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: [External] Re: [DIPY] Re: DTI and DKI processing This message was sent from a non-IU address. Please exercise caution when clicking links or opening attachments from external sources. Hi Suren, These warnings usually happen when your mask happens to include some voxels that are not in the brain (e.g., from the skull or the background surrounding the brain). Nothing to worry about, in almost all cases. Cheers, Ariel On Wed, Nov 24, 2021 at 5:36 PM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hello Dipy Community, I have been running DKI fitting using ‘restore’ method. However, I get this message. I am waiting for your favorable response as soon as possible. It says something about the line 1891 in dti.py. https://github.com/dipy/dipy/blob/master/dipy/reconst/dti.py Many Thanks, Suren [Text Description automatically generated] From: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Date: Wednesday, November 10, 2021 at 1:23 PM To: dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: DTI and DKI processing Hello Dipy Developers, My name is Suren. I am doing group-wise analysis of Alzheimer’s disease and healthy controls normal brain. I could successfully run DTI and DKI analysis. I did all other preprocessing steps: 1. Denoising (dwidenoise from MRtrix3) 2. Gibb’s artifact removal (from MRtrix3) 3. Eddy current correction and motion artifact removal (affine registration) Yet, I am wondering if we need data normalization and data regularization before performing DTI and DKI analysis. Also, I am wondering if we need to remove outliers before conducting both. I am using 3 shell dMRI data acquired in 9.4 Tesla MRI Scanner. I am waiting for your favorable response. Thank you very much. Best Regards, Suren Indiana University _______________________________________________ DIPY mailing list -- dipy@python.org<mailto:dipy@python.org> To unsubscribe send an email to dipy-leave@python.org<mailto:dipy-leave@python.org> https://mail.python.org/mailman3/lists/dipy.python.org/ Member address: arokem@gmail.com<mailto:arokem@gmail.com> _______________________________________________ DIPY mailing list -- dipy@python.org<mailto:dipy@python.org> To unsubscribe send an email to dipy-leave@python.org<mailto:dipy-leave@python.org> https://mail.python.org/mailman3/lists/dipy.python.org/ Member address: shreyasfadnavis@gmail.com<mailto:shreyasfadnavis@gmail.com>
Hi Surendra, My guess is that this may be an issue with DKI modelling itself because DKI is not supposed to be used on data that has bvals >= 3000. I still want to try it out on my end. Would it be possible to share NIFTI, BVEC and BVAL files with me (only -- will not share publicly)? I can test it out for you! Thanks, Shreyas On Thu, Nov 25, 2021 at 12:23 AM Maharjan, Surendra <smaharj@iu.edu> wrote:
Hi Surendra, I don't think the degeneracy in the fitting of DKI is stemming from Patch2Self but is from the DKI model itself. I suggest a possible workaround -- Use Patch2Self + Mean Signal DKI The Mean Signal Kurtosis can be obtained with no degeneracies as seen below: [cid:5894dc89-34e8-455d-a9d7-3c7fcd160be2] The tutorial for MSDKI can be found here: https://dipy.org/documentation/1.4.1./examples_built/reconst_msdki/#example-... The relevant math and details can be found in the following papers: 1. https://onlinelibrary.wiley.com/doi/pdf/10.1002/mrm.27606 2. https://www.frontiersin.org/articles/10.3389/fnhum.2021.675433/full Technically, MSK and MK should give you the same information ?? Thanks, Shreyas PS: Just so you know, DKI needs bvals < 3000. As per Ariel's suggestion, do try out using only those gradient directions with bvals < 3000. ________________________________ From: Ariel Rokem <arokem@uw.edu> Sent: Thursday, November 25, 2021 12:51 AM To: Shreyas Fadnavis <shreyasfadnavis@gmail.com> Cc: Maharjan, Surendra <smaharj@iu.edu>; dipy@python.org <dipy@python.org> Subject: [DIPY] Re: [External] Re: Re: DTI and DKI processing On Wed, Nov 24, 2021 at 9:31 PM Shreyas Fadnavis <shreyasfadnavis@gmail.com<mailto:shreyasfadnavis@gmail.com>> wrote: Hi Surendra, My guess is that this may be an issue with DKI modelling itself because DKI is not supposed to be used on data that has bvals >= 3000. That's a really good point! You might want to try this restricting to only b <= 3000, just to see whether that solves the problem. Ariel I still want to try it out on my end. Would it be possible to share NIFTI, BVEC and BVAL files with me (only -- will not share publicly)? I can test it out for you! Thanks, Shreyas On Thu, Nov 25, 2021 at 12:23 AM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hello Shreyas, I used like this: denoise = patch2self(data, bvals, model='ols', shift_intensity=True, clip_negative_vals=False, b0_threshold=100, b0_denoising=False) I tried b0_denoising both True and False, still gives me the black areas in the corpus callosum. Fitting is as follows: dki_model = DiffusionKurtosisModel(gtab, fit_method='WLS') dki_fit = dki_model.fit(data, mask=mask) dki_mk = dki_fit.mk<http://dki_fit.mk>() Many Thanks, Suren From: Shreyas Fadnavis <shreyasfadnavis@gmail.com<mailto:shreyasfadnavis@gmail.com>> Date: Thursday, November 25, 2021 at 12:17 AM To: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Cc: Ariel Rokem <arokem@uw.edu<mailto:arokem@uw.edu>>, dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: Re: [DIPY] Re: [External] Re: Re: DTI and DKI processing Hi Surendra, I don't think you should need to tune the alpha parameter in Patch2Self. Make sure that you are using DIPY 1.4.1. In DIPY >= 1.4.1, the default model is OLS, which will not require using alpha. If you still run into this issue, I am happy to help. I have worked on this issue in the past :) Thanks, Shreyas On Wed, Nov 24, 2021 at 11:18 PM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hello Ariel, Yes, I am using the default settings of patch2self. May I know the optimal value for alpha? Many Thanks, Suren From: Ariel Rokem <arokem@uw.edu<mailto:arokem@uw.edu>> Date: Wednesday, November 24, 2021 at 10:55 PM To: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Cc: dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: Re: [External] Re: [DIPY] Re: DTI and DKI processing If you aren't already, I'd recommend using Patch2Self, but make sure you are using the settings described in that issue (or a recent version of DIPY, that includes the fixes described therein). Ariel On Wed, Nov 24, 2021 at 7:51 PM Ariel Rokem <arokem@uw.edu<mailto:arokem@uw.edu>> wrote: Hi Suren, These are usually issues related to noise in the data. Remind me again: What are you doing to denoise this data? We saw similar things with some settings of Patch2Self, but that was resolved in https://github.com/dipy/dipy/issues/2334. Cheers, Ariel On Wed, Nov 24, 2021 at 7:31 PM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hi Ariel, Many Thanks for your valuable response. I am wondering why there’s null areas in the corpus callosum in the mice brain here in this image. This is the DKI- Mean Kurtosis Map. I used ‘weighted least square’, ‘ordinary least square’, ‘restore’, ‘non-linear least square’. All the fit methods have similar null signal areas in the corpus callosum. However, Mean Diffusivity (MD) is fine. The b-values from 0 to 7000, multi-shell data with 0.1 voxel size, multi-shot EPI sequence. [cid:17d558fb0554cff311] MK [A picture containing clothing Description automatically generated] MD Many Thanks, Suren From: Ariel Rokem <arokem@uw.edu<mailto:arokem@uw.edu>> Date: Wednesday, November 24, 2021 at 10:08 PM To: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Cc: dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: [External] Re: [DIPY] Re: DTI and DKI processing This message was sent from a non-IU address. Please exercise caution when clicking links or opening attachments from external sources. Hi Suren, These warnings usually happen when your mask happens to include some voxels that are not in the brain (e.g., from the skull or the background surrounding the brain). Nothing to worry about, in almost all cases. Cheers, Ariel On Wed, Nov 24, 2021 at 5:36 PM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hello Dipy Community, I have been running DKI fitting using ‘restore’ method. However, I get this message. I am waiting for your favorable response as soon as possible. It says something about the line 1891 in dti.py. https://github.com/dipy/dipy/blob/master/dipy/reconst/dti.py Many Thanks, Suren [Text Description automatically generated] From: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Date: Wednesday, November 10, 2021 at 1:23 PM To: dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: DTI and DKI processing Hello Dipy Developers, My name is Suren. I am doing group-wise analysis of Alzheimer’s disease and healthy controls normal brain. I could successfully run DTI and DKI analysis. I did all other preprocessing steps: 1. Denoising (dwidenoise from MRtrix3) 2. Gibb’s artifact removal (from MRtrix3) 3. Eddy current correction and motion artifact removal (affine registration) Yet, I am wondering if we need data normalization and data regularization before performing DTI and DKI analysis. Also, I am wondering if we need to remove outliers before conducting both. I am using 3 shell dMRI data acquired in 9.4 Tesla MRI Scanner. I am waiting for your favorable response. Thank you very much. Best Regards, Suren Indiana University _______________________________________________ DIPY mailing list -- dipy@python.org<mailto:dipy@python.org> To unsubscribe send an email to dipy-leave@python.org<mailto:dipy-leave@python.org> https://mail.python.org/mailman3/lists/dipy.python.org/ Member address: arokem@gmail.com<mailto:arokem@gmail.com> _______________________________________________ DIPY mailing list -- dipy@python.org<mailto:dipy@python.org> To unsubscribe send an email to dipy-leave@python.org<mailto:dipy-leave@python.org> https://mail.python.org/mailman3/lists/dipy.python.org/ Member address: shreyasfadnavis@gmail.com<mailto:shreyasfadnavis@gmail.com>
Hello Shreyas, Many Many Thanksgiving. Yes, the mean signal kurtosis (MSK) image looks nice and there are no null signal areas in CC. Yet, I am wondering how I can calculate MK, AK and RK from MSK (msdki) model. I really appreciate your time and effort. Best, Suren From: Fadnavis, Shreyas Sanjeev <shfadn@iu.edu> Date: Thursday, November 25, 2021 at 6:49 AM To: Ariel Rokem <arokem@uw.edu>, Shreyas Fadnavis <shreyasfadnavis@gmail.com>, Maharjan, Surendra <smaharj@iu.edu> Cc: dipy@python.org <dipy@python.org> Subject: Re: [DIPY] Re: [External] Re: Re: DTI and DKI processing Hi Surendra, I don't think the degeneracy in the fitting of DKI is stemming from Patch2Self but is from the DKI model itself. I suggest a possible workaround -- Use Patch2Self + Mean Signal DKI The Mean Signal Kurtosis can be obtained with no degeneracies as seen below: [cid:image001.png@01D7E1E4.38E19560] The tutorial for MSDKI can be found here: https://dipy.org/documentation/1.4.1./examples_built/reconst_msdki/#example-... The relevant math and details can be found in the following papers: 1. https://onlinelibrary.wiley.com/doi/pdf/10.1002/mrm.27606 2. https://www.frontiersin.org/articles/10.3389/fnhum.2021.675433/full Technically, MSK and MK should give you the same information 🙂 Thanks, Shreyas PS: Just so you know, DKI needs bvals < 3000. As per Ariel's suggestion, do try out using only those gradient directions with bvals < 3000. ________________________________ From: Ariel Rokem <arokem@uw.edu> Sent: Thursday, November 25, 2021 12:51 AM To: Shreyas Fadnavis <shreyasfadnavis@gmail.com> Cc: Maharjan, Surendra <smaharj@iu.edu>; dipy@python.org <dipy@python.org> Subject: [DIPY] Re: [External] Re: Re: DTI and DKI processing On Wed, Nov 24, 2021 at 9:31 PM Shreyas Fadnavis <shreyasfadnavis@gmail.com<mailto:shreyasfadnavis@gmail.com>> wrote: Hi Surendra, My guess is that this may be an issue with DKI modelling itself because DKI is not supposed to be used on data that has bvals >= 3000. That's a really good point! You might want to try this restricting to only b <= 3000, just to see whether that solves the problem. Ariel I still want to try it out on my end. Would it be possible to share NIFTI, BVEC and BVAL files with me (only -- will not share publicly)? I can test it out for you! Thanks, Shreyas On Thu, Nov 25, 2021 at 12:23 AM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hello Shreyas, I used like this: denoise = patch2self(data, bvals, model='ols', shift_intensity=True, clip_negative_vals=False, b0_threshold=100, b0_denoising=False) I tried b0_denoising both True and False, still gives me the black areas in the corpus callosum. Fitting is as follows: dki_model = DiffusionKurtosisModel(gtab, fit_method='WLS') dki_fit = dki_model.fit(data, mask=mask) dki_mk = dki_fit.mk<http://dki_fit.mk>() Many Thanks, Suren From: Shreyas Fadnavis <shreyasfadnavis@gmail.com<mailto:shreyasfadnavis@gmail.com>> Date: Thursday, November 25, 2021 at 12:17 AM To: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Cc: Ariel Rokem <arokem@uw.edu<mailto:arokem@uw.edu>>, dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: Re: [DIPY] Re: [External] Re: Re: DTI and DKI processing Hi Surendra, I don't think you should need to tune the alpha parameter in Patch2Self. Make sure that you are using DIPY 1.4.1. In DIPY >= 1.4.1, the default model is OLS, which will not require using alpha. If you still run into this issue, I am happy to help. I have worked on this issue in the past :) Thanks, Shreyas On Wed, Nov 24, 2021 at 11:18 PM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hello Ariel, Yes, I am using the default settings of patch2self. May I know the optimal value for alpha? Many Thanks, Suren From: Ariel Rokem <arokem@uw.edu<mailto:arokem@uw.edu>> Date: Wednesday, November 24, 2021 at 10:55 PM To: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Cc: dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: Re: [External] Re: [DIPY] Re: DTI and DKI processing If you aren't already, I'd recommend using Patch2Self, but make sure you are using the settings described in that issue (or a recent version of DIPY, that includes the fixes described therein). Ariel On Wed, Nov 24, 2021 at 7:51 PM Ariel Rokem <arokem@uw.edu<mailto:arokem@uw.edu>> wrote: Hi Suren, These are usually issues related to noise in the data. Remind me again: What are you doing to denoise this data? We saw similar things with some settings of Patch2Self, but that was resolved in https://github.com/dipy/dipy/issues/2334. Cheers, Ariel On Wed, Nov 24, 2021 at 7:31 PM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hi Ariel, Many Thanks for your valuable response. I am wondering why there’s null areas in the corpus callosum in the mice brain here in this image. This is the DKI- Mean Kurtosis Map. I used ‘weighted least square’, ‘ordinary least square’, ‘restore’, ‘non-linear least square’. All the fit methods have similar null signal areas in the corpus callosum. However, Mean Diffusivity (MD) is fine. The b-values from 0 to 7000, multi-shell data with 0.1 voxel size, multi-shot EPI sequence. [cid:image002.png@01D7E1E4.38E19560] MK [A picture containing clothing Description automatically generated] MD Many Thanks, Suren From: Ariel Rokem <arokem@uw.edu<mailto:arokem@uw.edu>> Date: Wednesday, November 24, 2021 at 10:08 PM To: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Cc: dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: [External] Re: [DIPY] Re: DTI and DKI processing This message was sent from a non-IU address. Please exercise caution when clicking links or opening attachments from external sources. Hi Suren, These warnings usually happen when your mask happens to include some voxels that are not in the brain (e.g., from the skull or the background surrounding the brain). Nothing to worry about, in almost all cases. Cheers, Ariel On Wed, Nov 24, 2021 at 5:36 PM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hello Dipy Community, I have been running DKI fitting using ‘restore’ method. However, I get this message. I am waiting for your favorable response as soon as possible. It says something about the line 1891 in dti.py. https://github.com/dipy/dipy/blob/master/dipy/reconst/dti.py Many Thanks, Suren [Text Description automatically generated] From: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Date: Wednesday, November 10, 2021 at 1:23 PM To: dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: DTI and DKI processing Hello Dipy Developers, My name is Suren. I am doing group-wise analysis of Alzheimer’s disease and healthy controls normal brain. I could successfully run DTI and DKI analysis. I did all other preprocessing steps: 1. Denoising (dwidenoise from MRtrix3) 2. Gibb’s artifact removal (from MRtrix3) 3. Eddy current correction and motion artifact removal (affine registration) Yet, I am wondering if we need data normalization and data regularization before performing DTI and DKI analysis. Also, I am wondering if we need to remove outliers before conducting both. I am using 3 shell dMRI data acquired in 9.4 Tesla MRI Scanner. I am waiting for your favorable response. Thank you very much. Best Regards, Suren Indiana University _______________________________________________ DIPY mailing list -- dipy@python.org<mailto:dipy@python.org> To unsubscribe send an email to dipy-leave@python.org<mailto:dipy-leave@python.org> https://mail.python.org/mailman3/lists/dipy.python.org/ Member address: arokem@gmail.com<mailto:arokem@gmail.com> _______________________________________________ DIPY mailing list -- dipy@python.org<mailto:dipy@python.org> To unsubscribe send an email to dipy-leave@python.org<mailto:dipy-leave@python.org> https://mail.python.org/mailman3/lists/dipy.python.org/ Member address: shreyasfadnavis@gmail.com<mailto:shreyasfadnavis@gmail.com>
Hi Suren, I find your processed DKI maps quite high quality. Is this an ex-vivo scan? Do you mind sharing your acquired protocol and scan time, the field strength, the coil info? Thank you! Best wishes Xiang On Thu, 25 Nov 2021 at 14:54, Maharjan, Surendra <smaharj@iu.edu> wrote:
Hello Xiang, Yes, these are ex-vivo scans. Field Strength: 9.4 T b-value: attached Voxel size: 0.1 Scan Protocol: Multishot EPI (EPI factor 8) No reversed phase encoding steps. Phase Encoding: L to R Scan Time:will tell you tomorrow Coil: Phase array coil 2x2 (without parallel imaging, SENSE) Many Thanks, Suren From: Xiang Feng <fengxiang.mr@gmail.com> Date: Thursday, November 25, 2021 at 12:25 AM To: Maharjan, Surendra <smaharj@iu.edu> Cc: Ariel Rokem <arokem@uw.edu>, dipy@python.org <dipy@python.org> Subject: Re: [DIPY] Re: [External] Re: Re: DTI and DKI processing Hi Suren, I find your processed DKI maps quite high quality. Is this an ex-vivo scan? Do you mind sharing your acquired protocol and scan time, the field strength, the coil info? Thank you! Best wishes Xiang On Thu, 25 Nov 2021 at 14:54, Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hi Ariel, Many Thanks for your valuable response. I am wondering why there’s null areas in the corpus callosum in the mice brain here in this image. This is the DKI- Mean Kurtosis Map. I used ‘weighted least square’, ‘ordinary least square’, ‘restore’, ‘non-linear least square’. All the fit methods have similar null signal areas in the corpus callosum. However, Mean Diffusivity (MD) is fine. The b-values from 0 to 7000, multi-shell data with 0.1 voxel size, multi-shot EPI sequence. [cid:image001.png@01D7E193.DCDD1F00] MK [A picture containing clothing Description automatically generated] MD Many Thanks, Suren From: Ariel Rokem <arokem@uw.edu<mailto:arokem@uw.edu>> Date: Wednesday, November 24, 2021 at 10:08 PM To: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Cc: dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: [External] Re: [DIPY] Re: DTI and DKI processing This message was sent from a non-IU address. Please exercise caution when clicking links or opening attachments from external sources. Hi Suren, These warnings usually happen when your mask happens to include some voxels that are not in the brain (e.g., from the skull or the background surrounding the brain). Nothing to worry about, in almost all cases. Cheers, Ariel On Wed, Nov 24, 2021 at 5:36 PM Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> wrote: Hello Dipy Community, I have been running DKI fitting using ‘restore’ method. However, I get this message. I am waiting for your favorable response as soon as possible. It says something about the line 1891 in dti.py. https://github.com/dipy/dipy/blob/master/dipy/reconst/dti.py Many Thanks, Suren [Text Description automatically generated] From: Maharjan, Surendra <smaharj@iu.edu<mailto:smaharj@iu.edu>> Date: Wednesday, November 10, 2021 at 1:23 PM To: dipy@python.org<mailto:dipy@python.org> <dipy@python.org<mailto:dipy@python.org>> Subject: DTI and DKI processing Hello Dipy Developers, My name is Suren. I am doing group-wise analysis of Alzheimer’s disease and healthy controls normal brain. I could successfully run DTI and DKI analysis. I did all other preprocessing steps: 1. Denoising (dwidenoise from MRtrix3) 2. Gibb’s artifact removal (from MRtrix3) 3. Eddy current correction and motion artifact removal (affine registration) Yet, I am wondering if we need data normalization and data regularization before performing DTI and DKI analysis. Also, I am wondering if we need to remove outliers before conducting both. I am using 3 shell dMRI data acquired in 9.4 Tesla MRI Scanner. I am waiting for your favorable response. Thank you very much. Best Regards, Suren Indiana University _______________________________________________ DIPY mailing list -- dipy@python.org<mailto:dipy@python.org> To unsubscribe send an email to dipy-leave@python.org<mailto:dipy-leave@python.org> https://mail.python.org/mailman3/lists/dipy.python.org/ Member address: arokem@gmail.com<mailto:arokem@gmail.com> _______________________________________________ DIPY mailing list -- dipy@python.org<mailto:dipy@python.org> To unsubscribe send an email to dipy-leave@python.org<mailto:dipy-leave@python.org> https://mail.python.org/mailman3/lists/dipy.python.org/ Member address: fengxiang.mr@gmail.com<mailto:fengxiang.mr@gmail.com>
participants (10)
-
Alex Crimi -
Ariel Rokem -
Cetin Karayumak, Suheyla -
Daniel Moyer -
Fadnavis, Shreyas Sanjeev -
Julio Villalón -
Maharjan, Surendra -
Rob Reid -
Shreyas Fadnavis -
Xiang Feng