Due 2/20

[ebridge2]

@ebridge2 Commitment:

• [x] Deliverable: methods.md for fft nuisance correction module, using 3 reference datasets (BNU1, DC1, HNU1). implement fMRI processing pipeline using FFT instead of component regression. show how it works on rs-fMRI (resting state) data with methods.md. This pipeline in THEORY will be better with task fMRI. simulations real data previous
• [x] Thursday: BMES Abstract. abstract

Stretch:

• [x] talk with Leo about a potential joint embedding approach to estimating latent covariance matrices. Due: notes on meeting with Leo.

@ewalke31

• [x] perform slice timing correction on more datasets and incorporate new metric (ask @ebridge2) to determine usefulness. Deliverable: ipython notebook with results

http://neurodatadesign.github.io/fngs/ewok_weekly/stc/Slice_Timing_Correction.html

@02agarwalt

• [x] Final registration benchmarking results: Link

[gkiar]

@ebridge2 I suggest making the following your goals for M and R:

• M: Methods.md on new post-processing of fMRI time series for task based analysis
• R: Decided which method to use and why (this will be helped by a conversation with Leo on joint embedding)

[ebridge2]

sounds good!

[ebridge2]

how's that look? yah I think this issue is definitely better for this week than the JE stuff; there were definitely a lot of moving parts with actually getting that done by next Monday. Good call :)

[gkiar]

Much better! One minor improvement: before all of that text for each task state in 5 words or less the deliverable. Thanks!!

[ebridge2]

Note: I apply frequency filter BEFORE doing remainder of nuisance correction so as to not have segmentation and nuisance correction related estimation errors impact the power of our frequency filtering. Least squares correction imposes a small error so if we were to apply that before fft there could be propogated error impacting the quality of our fft, or at least that's the logic I am going with.

[ebridge2]

Things to note: highpass with nuisance does a LOT better. however, it seems as though I cvould call it better (see real data top 2 plots; there is something weird going on in the high pass filtering where it is leaving the first timepoint untouched) so I am obviously doing something slightly wrong in that regard. overall seems promising though.

[gkiar]

• @ebridge2:
  • again I think you took an item off this list? You definitely had more bullets before. From now on I'm going to screenshot your issue at the start of the week and if it doesn't match and you don't tell me you've pivoted, you get 0 --- I'm trying to make communication consistent and hold you accountable to what you commit to.
• @ewalke31:
  • can you provide some description of your result? I want to know whether or not you think it is positive and what you think the next step is.
• @02agarwalt:
  • In future I'd also like to see some of your code performing this analysis, or all discriminability values plotted on the same axes, but this is quite dope. What do you think is the next step?

[ebridge2]

the issue was the original thursday deliverable, which i believe we agreed would be moved to this week instead of last (originally, it was something about talking with leo, and then reading a paper and deciding a strategy w him for interpretting task data) so i put it here as the second in my todo list. should i just leave it when i pivot next time?

[02agarwalt]

Summary:

My goal this week was to evaluate the performance of 3 registration combos:
1) FLIRT + LDDMM
2) NDRegAffine + LDDMM
3) NDRegAffine + FNIRT

The first 2 combos were dreadful for discriminability because the only way to run them realistically was to heavily downsample the data. The 3rd combo was a failure entirely because the transforms produced by NDReg are not compatible with standard fMRI registration techniques because NDReg doesn't register the anatomical image to the reference brain space. Kwame hasn't been able to provide a solution.

NDReg is clearly not suitable for fMRI processing. The best registration combo is FLIRT + FNIRT, so I think we should keep them. I'm not sure what to pursue next since this "chapter" of registration seems to be complete. Perhaps something involving epi_reg?

[ebridge2]

Summary:

My goal for this week was to evaluate the viability of a combined approach of Quadratic regression and frequency filtering to remove low frequency, slow drifts from our timeseries. This approach theoretically will be more favorable than compcor (our current strategy) particularly for task fMRI, as it is hypothesized that correcting for best-fit components from the white matter in gray matter can remove signal related to the hemodynamic response. Our simulations show that our algorithms generally work in the cases we expect (low frequency drift added, quadratic function added, particularly when the stimulus-related sinusoid has a frequency much greater than our cutoff and the noise is far below our cutoff). We can also see that when the noise is close to our cutoff, our model will perform relatively poorly, which shouldn't be much of a problem however as generally defining a cutoff around 0.01 Hz is relatively generous as low-frequency drift is often in the 0.001 - 0.005 Hz range. In our real data, we can see first and foremost that our first time point is relatively messed up, which is the result of our naive highpass filter not being able to compare the first time point to anything. As we can see, our new nuisance correction strategy unanimously improves our discriminability; particularly, it significantly improves our spectral discriminability (although our ranked correlational discriminability improves as well quite significantly).

Future investigations will hinge on looking into better highpass techniques (the current was a very naive algorithm consisting of fft -> cutoff -> ifft) such as applying specific box filters or butterworth filters, and seeing if we can improve the drift particularly in the early timesteps (note that aside from the first timestep being completely bogus it looks like a low-frequency drift remains in the first 50 timesteps of relatively low amplitude, but it is still present regardless). For this week, I have interviews Tuesday - Friday, but I will collect the existing discriminability code for spectral and correlational computation into a R package, and include a tutorial similar to the markdowns we have been working on for the last few weeks (a methods_markdown that basically is the same as the spectral_vs_corr_discr.html page, but shows in the context of this package itself). Also I am meeting with Leo in the AM and will have notes from our meeting that could be a Thursday deliverable.

[gkiar]

@ebridge2 yes leave it next time, please. As we discussed offline :)

@02agarwalt cool, so now let's look at epi_reg for fmri to t1/mprage registration. This will replace the aforementioned flirt only, not the flirt prior to running fnirt for t1 to mni registration.

@ebridge2 cool. reversing the time series will resolve first-timepoint issues, also applying a proper filter via convolution and some premade package may have considered t=0 filtering already and done some clever fix. GL on interviews!