Introduction

¹ is software for biomedical image analysis with a focus on registration, segmentation, geometric quantification and statistics. Statistical methods are available in ANTsR ² which tightly integrates ANTs with the R statistical computing language. This document briefly highlights ANTs features.

Provenance and Testing

Much core functionality in ANTs lives in ITK, a project to which we contribute regularly. This core is tested on many platforms via ctest and the results are on the ITK dashboard³. ANTs is also tested with every commit via codeship ⁴ and Travis ⁵. Many (not all) ANTs programs support programName --version. Our github commit hashes give the best way to track code versions by identifying core and dependency versions. See the ANTs website for current testing results and installation instructions.

In conjunction with other analysis systems

ANTs and ITK work synergistically within a well-defined I/O and world coordinate system. Using FSL, SPM or other pre-processing in conjunction with ITK and ANTs must be done with extreme care. The physical spaces (or interpretation of image headers) may not be the same in different systems. Such inconsistencies may lead to severe misinterpretation of results.

Data types

Registration

The most well-known components of ANTs relate to mapping between coordinate systems ¹⁰. Briefly, these maps may relate to:

• statistical image similarity: image difference, correlation or mutual information

• translation, rotation, affine, B-spline or various diffeomorphic models ¹¹

• spatial or spatiotemporal maps

• dense or sparse transformation models, the latter being similar to SIFT or HOGG

• applying composite transformations to images, labels or point sets.

These components all exist within an integrated framework which is nearly always capable of incorporating many features or applying to multiple different modalities.¹²

Quick start

ANTsR gives some quick registration options. One can achieve similar performance with antsRegistrationSyNQuick.sh.

fi <- antsImageRead(getANTsRData("r16") )
mi <- antsImageRead(getANTsRData("r64") )
mytxr <- antsRegistration(fixed=fi, moving=mi, typeofTransform = c('Rigid') )
mywarpedimager <- antsApplyTransforms( fixed=fi, moving=mi,
  transformlist=mytxr$fwdtransforms )
mytx <- antsRegistration(fixed=fi, moving=mi, typeofTransform = c('SyN') )
mywarpedimage <- antsApplyTransforms( fixed=fi, moving=mi,
  transformlist=mytx$fwdtransforms )

invisible( plot(fi, mi %>% iMath("Canny",1,5,12) ) )

initial relative image positions

invisible( plot(fi, mywarpedimager %>% iMath("Canny",1,5,12) ) )

rigidly aligned image positions

invisible( plot(fi, mywarpedimage %>% iMath("Canny",1,5,12) ) )

quick deformable registration

Segmentation, geometry and labeling

This category of methods relates to labeling and quantifying images. These methods may require data of reasonable quality to perform well.

N4

Nick’s N3 ¹⁵ improves the original N3 inhomogeneity correction method.

ficorrupt = as.antsImage( as.array( fi ) * 1:length(fi[fi>=0]) )
fifixed = n4BiasFieldCorrection( ficorrupt )

invisible( plot( ficorrupt ) )

corrupted image

invisible( plot( fifixed ) )

n4 repair

Atropos

Expecation maximization segmentation with a variety of likelihood models and initialization strategies. Incorporates multiple modalities and allows control of prior strength. The finite mixture modeling (FMM) segmentation approach is the most popular. Prior constraints include the specification of a prior label image, prior probability images (one for each class), and/or an MRF prior to enforce spatial smoothing of the labels.

img <- antsImageRead( getANTsRData("r16")  )
mask <- getMask(img)
segs1 <- kmeansSegmentation( img, 3, mask )

# Use probabilities from k-means seg as priors
feats <- list(img, iMath(img,"Laplacian"), iMath(img,"Grad") )
segs2 <- atropos( d = 2, a = feats, m = '[0.2,1x1]',
   c = '[2,0]',  i = segs1$probabilityimages, x = mask )

invisible( plot( fi, segs1$segmentation ) )

kmeans result

invisible( plot( fi, segs2$segmentation ) )

prior probability result with 2 feature channels

Weingarten map-based surface curvature and area

The shape operator provides a beautiful way to compute the mean or Gaussian curvature (or related values) in any three-dimensional image ¹⁶. The ANTs program that implements this is called SurfaceCurvature.

fi <- antsImageRead( getANTsRData( "mni" ) )
fiseg = kmeansSegmentation( fi, 3 )
fik <- weingartenImageCurvature( fi )

invisible( plot(  fik, axis=3 ) )

shape operator in 3D

fisulc = antsImageClone( fik ) * 0
selector = ( fiseg$segmentation == 2 & fik < 0 )
fisulc[  selector ] = fik[ selector ]
invisible( plot(  fi, fisulc, axis=3 ) )

shape operator selects sulci

Cortical thickness pipeline

You can use antsCorticalThickness.sh to reproduce our lifespan thickness analysis from the Freesurfer evaluation paper ¹⁷.

img<-antsImageRead( getANTsRData("r16") ,2)
mask<-getMask( img )
segs<-kmeansSegmentation( img, k=3, kmask = mask)
thk<-kellyKapowski( s=segs$segmentation, g=segs$probabilityimages[[2]],
       w=segs$probabilityimages[[3]],its=45,r=0.5,m=1 )

invisible( plot(  img, thk, window.overlay = c(0.1,max(thk)) ) )

thickness calculation

Multiple modality pipeline

You can use our neurobattery to reproduce our adolescent multiple modality analysis from the paper ¹⁸. The neurobattery is online and administered by Jeffrey T. Duda ¹⁹.

Brain mapping in the presence of lesions or other obstructions

Our fully automated registration and segmentation approach for the lesioned or occluded brain uses machine learning to identify and down-weight missing data. One may also use pre-identified inclusive masks to focus registration on healthy tissue ²⁰.

Statistical pipelines

ANTs-based statistical pipelines are best accessed through ANTsR. We use these pipelines for:

• estimating cerebral blood flow with robust and data-driven methods

• performing structural and functional connectivity measurements

• employing biologically-regularized dimensionality reduction for hypothesis testing

• implementing exploratory multivariate analysis tools such as sparse canonical correlation analysis for neuroimaging ²¹

• performing prediction studies ²²

• associating modalities ²³

and more basic studies such as mass univariate regression.

Overview of ANTs programs

ANTs executables come in either binary ( compiled from C++ ) or script form ( usually bash or R ). Here, we summarize the relevant programs.