New CompCor interface

[oesteban]

Create a new interface (under algorithm.misc ?) for the calculation of the Component Based Noise Correction Method (CompCor) on r-fMRI.

The code is already here in nipype:

• https://github.com/nipy/nipype/blob/master/nipype/workflows/rsfmri/fsl/resting.py
• https://github.com/nipy/nipype/blob/c389d5d4c0f1eb5ff321de5553f815d1fb6fcc47/examples/rsfmri_vol_surface_preprocessing_nipy.py

So this issue would include updating both workflow and example to use the new interface.

• implement
  • CompCor
    • remove regression residuals
  • tCompCor
    • threshold param
  • aCompcor alias
• use TSNR
  • refactor TSNR
  • move TSNR back into misc.py
  • use in compcor.py
• use interface in https://github.com/nipy/nipype/blob/master/nipype/workflows/rsfmri/fsl/resting.py
• use interface in https://github.com/nipy/nipype/blob/c389d5d4c0f1eb5ff321de5553f815d1fb6fcc47/examples/rsfmri_vol_surface_preprocessing_nipy.py
  • refactor
  • runthrough
• fix ups
  • doctest
  • use tempfile in test_tsnr.py
  • explicit relative imports
  • boolean flag for regress_poly

[chrisgorgo]

Big thumbs up!

On Aug 29, 2016 1:47 PM, "Oscar Esteban" notifications@github.com wrote:

Create a new interface (under algorithm.misc ?) for the calculation of the Component
Based Noise Correction Method (CompCor)
http://dx.doi.org/10.1016%2Fj.neuroimage.2007.04.042 on r-fMRI.

The code is already here in nipype:

• https://github.com/nipy/nipype/blob/master/nipype/
  workflows/rsfmri/fsl/resting.py
  https://github.com/nipy/nipype/blob/master/nipype/workflows/rsfmri/fsl/resting.py
• https://github.com/nipy/nipype/blob/c389d5d4c0f1eb5ff321de5553f815
  d1fb6fcc47/examples/rsfmri_vol_surface_preprocessing_nipy.py
  https://github.com/nipy/nipype/blob/c389d5d4c0f1eb5ff321de5553f815d1fb6fcc47/examples/rsfmri_vol_surface_preprocessing_nipy.py

So this issue would include updating both workflow and example to use the
new interface.

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
#1594, or mute the thread
https://github.com/notifications/unsubscribe-auth/AAOkp7nwsvjUgNL5oaNtJqyDBuWOLGBEks5qkxtKgaJpZM4Jvt6N
.

[chrisgorgo]

It would be a "pure python" interface similar to https://github.com/nipy/nipype/blob/master/nipype/algorithms/icc.py#L25

[satra]

the function implemented in the resting workflow assumes that something else computes the voxel mask over which the PCA is implemented. this allows the a and t pieces to be taken care of by other functions/workflows. for example, fast is used to get the tissue classes. i would keep that separation for this interface. you can then build a sequence of nodes for aCompCor and tCompCor.

[berleant]

are you saying that computing the t mask (top 2% tSTDs) is already implemented somewhere? That would be great

[berleant]

In the paper, after the components are computed, they are applied using a GLM analysis with an AR(1) model for the additive noise component. There are many options in nipype for doing a GLM. The examples above use fsl.FilterRegressor and fsl.GLM, but there is also freesurfer.GLMFit, which has a parameter for using AR1. Which interface is preferred? There is also the option of leaving it up to the user.

[chrisgorgo]

I would not incorporate applying of the components into this interface -
this step usually combines regressors coming from multiple sources in
addition to CompCor (motion, task design, physiological recordings etc). If
you need this for testing I would go with FSL or something from statmodels.

On Sun, Sep 4, 2016 at 4:52 PM, Shoshana Berleant notifications@github.com
wrote:

In the paper, after the components are computed, they are applied using a
GLM analysis with an AR(1) model for the additive noise component. There
are many options in nipype for doing a GLM. The examples above use
fsl.FilterRegressor and fsl.GLM, but there is also freesurfer.GLMFit, which
has a parameter for using AR1. Which interface is preferred?

—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub
#1594 (comment), or mute
the thread
https://github.com/notifications/unsubscribe-auth/AAOkp3ieWZjDTYCQXRFQcvZOKz2jERJAks5qm1nIgaJpZM4Jvt6N
.

[berleant]

Oh good, I was leaning that way myself

On Sun, Sep 4, 2016, 5:22 PM Chris Filo Gorgolewski <
notifications@github.com> wrote:

I would not incorporate applying of the components into this interface -
this step usually combines regressors coming from multiple sources in
addition to CompCor (motion, task design, physiological recordings etc). If
you need this for testing I would go with FSL or something from statmodels.

On Sun, Sep 4, 2016 at 4:52 PM, Shoshana Berleant <
notifications@github.com>
wrote:

In the paper, after the components are computed, they are applied using a
GLM analysis with an AR(1) model for the additive noise component. There
are many options in nipype for doing a GLM. The examples above use
fsl.FilterRegressor and fsl.GLM, but there is also freesurfer.GLMFit,
which
has a parameter for using AR1. Which interface is preferred?

—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub
#1594 (comment), or
mute
the thread
<
https://github.com/notifications/unsubscribe-auth/AAOkp3ieWZjDTYCQXRFQcvZOKz2jERJAks5qm1nIgaJpZM4Jvt6N

.

—
You are receiving this because you were assigned.
Reply to this email directly, view it on GitHub
#1594 (comment), or mute
the thread
https://github.com/notifications/unsubscribe-auth/ABVoRLyu0d7-KtxzjOMl6eQMCu1AKPKBks5qm2DPgaJpZM4Jvt6N
.

[berleant]

In the paper, they find the tSTD "after the removal of low-frequency nuisance terms (e.g., linear and quadratic drift)."

Can I assume, if the input is an already-aligned image, that this processing has already been done? If not, how does one determine the "low-frequency nuisance terms"?

[satra]

@shoshber - in the resting state workflow, the TSNR node takes care of this

[berleant]

ok, I will assume it is already done