Improving skullstripping and nonlinear warping in AFNI: sswarper2

1691 

Abstract Submission 

Paul Taylor¹, Richard Reynolds¹, Daniel Glen²

¹National Institute of Mental Health, Bethesda, MD, ²NIMH, Bethesda, MD

Paul Taylor  
National Institute of Mental Health
Bethesda, MD

Richard Reynolds  
National Institute of Mental Health
Bethesda, MD

Daniel Glen  
NIMH
Bethesda, MD

Nonlinear alignment (or registration, or spatial "normalization") is a common processing step in FMRI analysis. While it is computationally expensive and can require on the order of hours to complete, it allows for individual datasets to be correspond across a common space. As more similar structures are brought into closer alignment, the accuracy and statistical power of the final results should increase.

However, no alignment algorithm is perfect. Mathematically, estimated "warps" between the source and reference datasets are generally constrained to be smooth, invertible and diffeomorphic. But even among healthy populations, anatomical variability exists (e.g., numbers of sulci/gyri can differ locally), and warps are typically unable to overcome such differences. Moreover, datasets can have distortion, low contrast, etc., possibly decreasing quality of alignment. As a result, algorithms can get stuck in local minima or produce distortions or other inaccuracies.

Here we describe "sswarper2," a new nonlinear alignment program in AFNI [1] that both skullstrips (SS) a T1w volume and aligns (warps) it to a reference dataset. We show its improvements to predecessor @SSwarper, which had similar dual roles.

The primary nonlinear alignment program in AFNI is 3dQwarp [2]. @SSwarper wraps around 3dQwarp and also includes removal of nonbrain material ("skullstripping"), since warping and skullstripping are each complicated processes that improve when the other is done well. If the skull+nonbrain region of a subject's anatomical dataset has been exactly removed, alignment to a template is much easier; or if a dataset has been well aligned to a template, one can use the latter to "punch away" the skull of the former. @SSwarper makes use of this by iterating between these steps with increasing accuracy, improving each. During alignment, it uses a local Pearson cost to drive early stages of alignment [3]. It also saves snapshots of initial and final stages of overlap, to provide quality control (QC) checks.

Similarly, sswarper2 iterates between (local Pearson) alignment to a template and skullstripping, but it does so in smaller steps, more finely interleaved with skull removal, increasing stability. Additionally, sswarper2 saves a detailed history of snapshots of intermediate processing steps, to facilitate any troubleshooting.

To compare overall robustness of the programs, we tested each on a set of 169 anatomical T1w datasets from 8 different sites from 3 continents, with a wide subject age range (8-70 yrs) [4], aligning to the MNI 2009c asymmetric template [5]. We compare individual and group-wide results both qualitatively and quantitatively.

In many cases, @SSwarper and sswarper2 yield quite similar results, providing accurate skullstripping and nonlinear registration to the MNI template. However, in a small number of cases, @SSwarper had inaccurate final skullstripping, resulting in localized misalignment. Fig 1 shows two examples of datasets in which the warped subject anatomy extends 2-3 mm outside the brain locally. In each case, sswarper2 provides more accurate alignment and skullstripping.

Fig 2A shows the mean of all warped datasets (each had been unifized to similar brightness values per tissue class) for each program. There is overall good alignment to the template in both cases. Fig 2B shows the voxelwise standard deviation across the group for each program. Again, overall patterns are similar, but sswarper2 results are uniformly more tightly aligned within the brain volume and at major tissue boundaries, while the effect of the small fraction of stretched @SSwarper results can be seen around the brain edge.

While both programs typically produce warped anatomicals that closely match the template structure, the new sswarper2 provides more robust alignment across a broader range of datasets. The sswarper2 results can also be integrated directly into afni_proc.py and other FMRI pipeline tools.

Classification and Predictive Modeling

Exploratory Modeling and Artifact Removal ¹

Methods Development ²

Computing

Data analysis

Design and Analysis

FUNCTIONAL MRI

Modeling

NORMAL HUMAN

Open Data

Open-Source Software

Workflows