nipy / dmriprep Goto Github PK

We want to produce an image of the DTI ellipsoids around the posterior corpus callosum in a horizontal section. If these are pointing orthogonal to the posterior callosum, rather than swinging back like we would expect, the b vectors need to be reoriented.

Might be useful to provide example images that tell you exactly what you should do (i.e., what you should provide as keyword arguments to dmriprep upon re-running it).

Right now, we have to drop into our docker image. In the long run, we'd also like to run this from the command line, without dropping in (see #72).

get_all_keys (or whatever it's called) should accept a subject keyword. If it's None (the default), download all subjects.

AssertionError: found 0 ap fieldmap files and we need just 1

Leftover issue from #38
Subjects from Site-CBIC have multiple anatomical directories, e.g.

• T1w_VNav
• T1w_VNavNorm
• T1w_HCP

We need to decide which one to use and filter the S3 keys based on that choice.

I think that typically an s3 bucket would have just one study organized as BIDS within it.

We need two pipelines: one with FSL eddy and one with FLIRT. FSL eddy is preferred so the CLI should determine whether the input dataset meets the requirements for eddy. There are two possible ways to satisfy the eddy requirements:

• A set of diffusion encoding directions that span the entire sphere and not just a half-sphere
• A blip-up-blip-down (phase encode reversed) acquisition

So if the input data has phase encoded files, then we can safely use the eddy-based pipeline. If not, we should determine in the bvecs all lie within one hemisphere. If so, fallback to FLIRT. If not, use eddy. To determine whether the bvecs all lie within any one arbitrary hemisphere, see this for a description of a deterministic O(n^3) algorithm to determine if the points are within a hemisphere.

Or check the hash of pre-existing files to confirm that they match what's on S3.

Plot the number of outliers per volume. This way the user can set a threshold to remove volumes with a lot of outliers

Write a manuscript

@arokem : we are dropping the volumes that have > threshold outlier slices. The threshold can be defined as a number or percentation of the total number of slices. but is it safe to assume that the 2nd axis is the slice dimension? I'm thinking it's not -- is there a way with nibel to get the index of the slice dimension?

https://github.com/nipy/dmriprep/blob/master/dmriprep/run.py#L312

for some odd reason img.get_fdata kills my python kernel, but the deprecated get_data does not. I'm not sure how to debug this because everything just dies!

suggested by @satra

Default should be to skip topup when there is not both AP and PA phase encodes

Package up Freesurfer to run with Cloudknot/AWS.

I already have the code here: https://github.com/uw-biomedical-ml/mri2mri_notebooks/blob/master/ck-freesurfer-segmentations.ipynb

There's code for that here: https://github.com/yeatmanlab/pyAFQ/blob/master/AFQ/registration.py

NBD

I'm relatively new to dwi preprocessing. Would adding the following steps (added the mrtrix commands) as optional steps before eddy be useful?

• denoising (dwidenoise)
• unringing (mrdegibbs)
• upsampling (mrresize)

Inputs: threshold

• if threshold is an int >= 1, treat as number of allowed outlier slices before we drop the whole volume.
• if threshold is a float between 0 and 1 (exclusive), treat it as the fraction of allowed outlier slices before we drop the whole volume.

QC pipeline should reject bad volumes and write out a QCed dataset (and bvals and bvecs) that has bad volumes (and correspondingm bvecs) removed. We will set suggested thresholds based on maximizing scan rescan reliability. But it would be AWESOME if the plots had sliders on the yaxis where the user could adjust this

Right now, we implicitly assume that there is just one run of dMRI in each session. But that doesn't have to be the case.

Basically wrap the ungodly docker command that you would have run so that all you would have to do is something like:

dmriprep_docker $BIDS_FOLDER

Another possible way to deal with EPI distortions is to collect two sets of DWI with two inverted phase-encode directions, but no field maps.

Does topup know how to deal with this as input? Should we add this as an option here?

the HBN fieldmap files aren't consistently in BIDS format -- for the files where we have to read the .json file to figure out the fieldmap direction, we need to save the file with _dir-AP and _dir-PA in the filename so that pybids can find the right files in the io module. Most of the Site-RU subjects don't have the right naming scheme (e.g. sub-NDARUM797TFA from RU)

i completely think dmriprep should focus on "standard" brain scans to start with. but will it be written in a way to allow for possible modification to cover other brain regions scanned with diffusion imaging. for example:

1. brainstem
2. spinal cord
3. partial FOV

For the QC registration plot add an axial slice and a coronal slice

This currently fails on a call to fslmerge that requires equishaped volumes.

Make public by default

There are a few mentions to preAFQ. Will need to be renamed to dmriprep.

Write tests

Make it a parameter (default=5)

so that its easy for people to download data and get started

It would be nice if this could (optionally) take in a T1 weighted image and align the dMRI data to this T1. Thus the dMRI data will be in the same space as the freesurfer segmentation

at least for development its a lot easier

• also make the environment called "dmriprep" instead of "preafq"

we need to add anatomical processing if freesurfer hasn't been run. Let's use the same one as fmriprep: https://github.com/poldracklab/fmriprep/blob/master/fmriprep/workflows/base.py#L409-L466

This should look just like other BIDS apps. Use the example app as a template.

Add the suggestions from https://neurostars.org/t/dmriprep-an-fmriprep-for-diffusion/2385/8 to the QC report

in the get_dataset function, we need to run mri_convert on the HBN data where the T1.nii.gz is in a weird spot

Use https://www.biorxiv.org/content/early/2018/07/24/306951 for inspiration.

Demonstrate the choices made as defaults are based on empirical evidence from test-retest data.

See: https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/eddy/UsersGuide#A--repol

from dmriprep.data import get_dataset
get_dataset('path/to/output/dir') # needs a progress bar in bash and also on notebook?

• also would be nice if pathlib made the folder for you if it doesn't already exist mkdir -p
• maybe remove Site-SI from output folder structure

Per conversation with @akeshavan, I'm putting this here for some measure of posterity.

Right now smriprep/fmriprep does not handle preprocessed Freesurfer for longitudinal data (using the Freesurfer longitudinal recon-all pipeline). See this issue ([https://github.com/nipreps/smriprep/issues/21]).

Implementation that allows for preprocessed longitudinal Freesurfer is something to consider if dmriprep is going to allow/require preprocessed data.

Get a small data set and run the whole thing end-to-end on it. Manipulate to test (for example) that wonky b-vectors can be detected, that it works with an odd number of slices, and so forth.

The output of dmriprep currently has:

In [7]: hdr = img.header                                                        

In [8]: hdr.get_dim_info()                                                      
Out[8]: (None, None, None)

We want to set this to be consistent with the input. See #32

The workflow should interrogate the host system to see if there is a GPGPU available. If so, use eddy_cuda instead of eddy_openmp. This is accomplished with eddy.inputs.use_cuda=True.

To determine if a cuda capable GPU is present, we could try

import numba.cuda
if numbda.cuda.gpus.current:
    eddy.inputs.use_cuda = True

or something like that.

a la MRI QC, so that we can provide QC information relative to a norm.

For the QC image of the RGB map it would be helpful to also have a small white line showing the direction vector of the PDD taken from the tensor

Two kinds of use-cases:

• Acquisition protocols that don't include T1w.
• Individual cases within a protocol that don't have that data (e.g., when the quality of the T1 is not good. This is a thing in the HBN dataset).

One option is to register to the FS average brain.

• in BIDS format
• cases without T1s and with T1s, without AP/PA and with

https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/eddyqc/UsersGuide