multi-res voxelwise neighborhood random forest segmentation learning

Represents multiscale feature images as a neighborhood and uses the features to build a random forest segmentation model from an image population

mrvnrfs(
  y,
  x,
  labelmasks,
  rad = NA,
  nsamples = 1,
  ntrees = 500,
  multiResSchedule = c(4, 2, 1),
  asFactors = TRUE,
  voxchunk = 50000,
  ...
)

Arguments

y	list of training labels. either an image or numeric value
x	a list of lists where each list contains feature images
labelmasks	a mask (or list of masks) used to define where the samples will come from. Note, two labels (e.g., GM and WM) will double the number of samples from each feature images. If the mask is binary, samples are selected randomly within 1 values.
rad	vector of dimensionality d define nhood radius
nsamples	(per subject to enter training)
ntrees	(for the random forest model)
multiResSchedule	an integer vector defining multi-res levels
asFactors	boolean - treat the y entries as factors
voxchunk	value of maximal voxels to predict at once. This value is used to split the prediction into smaller chunks such that memory requirements do not become too big
...	arguments to pass to vwnrfs

Value

list a 4-list with the rf model, training vector, feature matrix and the random mask

Author

Avants BB, Tustison NJ, Pustina D

Examples

mask<-makeImage( c(10,10), 0 )
mask[ 3:6, 3:6 ]<-1
mask[ 5, 5:6]<-2
ilist<-list()
lablist<-list()
inds<-1:5
scl<-0.33 # a noise parameter
for ( predtype in c("label","scalar") )
{
for ( i in inds ) {
  img<-antsImageClone(mask)
  imgb<-antsImageClone(mask)
  limg<-antsImageClone(mask)
  if ( predtype == "label") {  # 4 class prediction
    img[ 3:6, 3:6 ]<-rnorm(16)*scl+(i %% 4)+scl*mean(rnorm(1))
    imgb[ 3:6, 3:6 ]<-rnorm(16)*scl+(i %% 4)+scl*mean(rnorm(1))
    limg[ 3:6, 3:6 ]<-(i %% 4)+1  # the label image is constant
    }
    if ( predtype == "scalar") {
      img[ 3:6, 3:6 ]<-rnorm(16,1)*scl*(i)+scl*mean(rnorm(1))
      imgb[ 3:6, 3:6 ]<-rnorm(16,1)*scl*(i)+scl*mean(rnorm(1))
      limg<-i^2.0  # a real outcome
      }
    ilist[[i]]<-list(img,imgb)  # two features
    lablist[[i]]<-limg
  }
rad<-rep( 1, 2 )
mr <- c(1.5,1)
rfm<-mrvnrfs( lablist , ilist, mask, rad=rad, multiResSchedule=mr,
     asFactors = (  predtype == "label" ) )
rfmresult<-mrvnrfs.predict( rfm$rflist,
     ilist, mask, rad=rad, asFactors=(  predtype == "label" ),
     multiResSchedule=mr )
if ( predtype == "scalar" )
  print( cor( unlist(lablist) , unlist( rfmresult$seg ) ) )
} # end predtype loop
#> randomForest 4.6-14
#> Type rfNews() to see new features/changes/bug fixes.
#> Warning: The response has five or fewer unique values.  Are you sure you want to do regression?
#> Warning: The response has five or fewer unique values.  Are you sure you want to do regression?
#> [1] 0.8239526