tsalo/complex-flow

Initial work on a complex-valued fMRI preprocessing workflow.

complex-flow

Initial work on a complex-valued fMRI preprocessing workflow.

Goals

• Processing of both magnitude and phase data for both single- and multi-echo data
• BIDS Derivatives-compatible outputs
• A user-friendly CLI
• Support for phase-based denoising within the workflow (when applicable), including phase regression
• Support for coil-level data
• Dynamic distortion correction, for multi-echo data, with DOCMA
• Support for field maps and single-band reference images
• Long-term, I want to move any workflows produced here into tools like SDCFlows and niworkflows

complex-flow's Issues

Estimate magnitude and phase error variance for phase regression

Per Curtis 2014, the fit of the phase regressor should be conditioned on the relative error levels in the measurements (magnitude and phase signal). A quote:

As a coarse measure, one can estimate the standard deviations σS and σφ of the magnitude and phase time series respectively, if the signal changes of interest can be factored out.

It looks like they did this in task data by performing a task-based regression using the convolved task signal and then estimating standard deviations from the residuals. With resting state data (as in Curtis 2014), they "removed" ostensible BOLD signal with a bandpass filter. I'm not a huge fan of incorporating task information in a preprocessing pipeline, and it's unclear to me if the regression is actually performed on the full data or just the residuals.

Strange edges in unwrapped EPI phase maps

After rescaling according to the FSL instructions and unwrapping with PRELUDE, I still see odd edges in the phase data, as below:

The edges are not stable across echoes, so the calculated phase-difference images end up with odd edges as well, as below:

This happens with both MB EPI and single-band reference images, but not with the field maps, so I wonder if it's something to do with the sequence or with the long echo times.

Test dwidenoise as part of preprocessing workflow

dwidenoise should be performed before any other preprocessing steps, and can be run on either magnitude-only or complex data. Some initial tests show minimal differences, at least for low-resolution data, on tedana components. Still, it may be worth pursuing within the context of a larger workflow.

One note- the phase data after the complex denoising seems to be unusable.

Denoising strategies to consider

1. Multi-echo ICA
   • See tedana.readthedocs.io.
   • Separation of BOLD from non-BOLD noise (e.g., motion).
   • Use tedana.
2. Phase regression
   • Improved spatial localization by suppressing macrovascular signal.
   • Use phaseprep.
3. Physiological trace regression
   • Presumably separation of neural BOLD signal and non-neural BOLD-based noise, but limited to (1) variations in heart rate and respiratory volumes and (2) carbon dioxide.
   • Use phys2denoise.
4. Dynamic global signal regression
   • See Tong, Hocke, and Frederick (2019).
   • Separation of neural BOLD signal and non-neural BOLD-based noise.
   • Likely sources of this non-neural BOLD-based noise include:
     • Variations in heart rate and respiratory volumes
     • Carbon dioxide
     • Vasomotion from oscillations in the vascular tone
     • Gastric oscillations
   • Use rapidtide.
5. Marchenko-Pastur PCA dimensionality reduction
   • See Adhikari et al. (2019).
   • Removal of thermal noise(?)
   • Per anecdotal evidence, dwidenoise can improve ME-ICA, including more subcortical regions in components with high-resolution fMRI data.
   • Also, the ENIGMA consortium uses it for fMRI data.
   • Use dwidenoise function in mrtrix3.
6. Filtering of motion parameters before regression or censoring
   • See Gratton et al. (2020).
   • Respiration-induced factitious head motion via B0 perturbations along the phase-encoding direction.

Test ROMEO for multi-echo unwrapping

A link to the binary: https://github.com/korbinian90/ROMEO/releases/tag/v1.4

Additional preprocessing steps to consider

1. Dynamic distortion correction
   • Requires accurate unwrapping across echoes.
   • Prior to motion correction, since it will correct distortion on a volume-by-volume basis.
2. Coregistration using single-band reference images averaged across echoes?

Correction for background field gradients in ME data

I think that the method described in Cohen Adad et al. (2012) could be used with complex multi-echo EPI data for more accurate T2* estimation.