jointLabelFusion.RdA multiple atlas voting scheme to customize labels for a new subject. This
function will also perform intensity fusion. It almost directly calls the
C++ in the ANTs executable so is much faster than other variants in ANTsR.
One may want to normalize image intensities for each input image before
passing to this function. If no labels are passed, we do intensity fusion.
Note on computation time: the underlying C++
is multithreaded. You can control the number of threads by setting the
environment variable ITK_GLOBAL_DEFAULT_NUMBER_OF_THREADS e.g. to use all or
some of your CPUs. This will improve performance substantially. For instance,
on a macbook pro from 2015, 8 cores improves speed by about 4x.
jointLabelFusion(
targetI,
targetIMask,
atlasList,
beta = 4,
rad = 2,
labelList = NULL,
rho = 0.01,
usecor = FALSE,
rSearch = 3,
nonnegative = FALSE,
maxLabelPlusOne = FALSE,
noZeroes = FALSE,
verbose = FALSE
)antsImage to be approximated
mask with value 1
list containing antsImages with intensity images
weight sharpness, default to 2
neighborhood radius, default to 2
optional list containing antsImages with segmentation labels
ridge penalty increases robustness to outliers but also makes image converge to average
employ correlation as local similarity
radius of search, default is 3
constrain weights to be non-negative
boolean this will add max label plus one to the non-zero parts of each label where the target mask is greater than one. NOTE: this will have a side effect of adding to the original label images that are passed to the program. It also guarantees that every position in the labels have some label, rather than none. Ie it guarantees to explicitly parcellate the input data.
boolean will zero out target mask regions that have any zero label. this prevents JLF from computing a solution in regions not covered by the initial library.
boolean
approximated image, segmentation and probabilities
set.seed(123)
ref <- ri(1)
ref <- resampleImage(ref, c(50, 50), 1, 0)
ref <- iMath(ref, "Normalize")
mi <- ri(2)
mi2 <- ri(3)
mi3 <- ri(4)
mi4 <- ri(5)
mi5 <- ri(6)
refmask <- getMask(ref)
refmask <- iMath(refmask, "ME", 2) # just to speed things up
ilist <- list(mi, mi2, mi3, mi4, mi5)
seglist <- list()
for (i in 1:length(ilist))
{
ilist[[i]] <- iMath(ilist[[i]], "Normalize")
mytx <- antsRegistration(
fixed = ref, moving = ilist[[i]],
typeofTransform = c("Affine"), verbose = TRUE
)
mywarpedimage <- antsApplyTransforms(
fixed = ref,
moving = ilist[[i]],
transformlist = mytx$fwdtransforms
)
ilist[[i]] <- mywarpedimage
seg <- thresholdImage(ilist[[i]], "Otsu", 3)
seglist[[i]] <- seg
}
r <- 2
pp <- jointLabelFusion(ref, refmask, ilist,
rSearch = 2,
labelList = seglist, rad = rep(r, length(dim(ref)))
)
#> Warning: converting NULL pointer to R NULL
#>
|
| | 0%
|
|======================= | 33%
|
|=============================================== | 67%
|
|======================================================================| 100%
pp2 <- jointLabelFusion(ref, refmask, ilist,
rSearch = 2,
labelList = seglist, rad = rep(r, length(dim(ref)))
)
#> Warning: converting NULL pointer to R NULL
#>
|
| | 0%
|
|======================= | 33%
|
|=============================================== | 67%
|
|======================================================================| 100%
testthat::expect_equal(pp2$segmentation, pp$segmentation)
pp <- jointLabelFusion(ref, refmask, ilist,
rSearch = 2,
rad = rep(r, length(dim(ref)))
)
#> Warning: converting NULL pointer to R NULL
if (FALSE) { # \dontrun{
ref <- antsImageRead(getANTsRData("ch2"))
n <- 50
ref <- resampleImage(ref, c(n, n, n), 1, 0)
ref <- iMath(ref, "Normalize")
refmask <- getMask(ref)
ilist <- list()
seglist <- list()
for (k in 1:5) {
mi <- antsImageClone(ref) + rnorm(n^3, 0, 0.1)
mykseg <- kmeansSegmentation(mi, 3, refmask)$segmentation
ilist[[k]] <- mi
seglist[[k]] <- mykseg
}
pp <- jointLabelFusion(ref, refmask, ilist,
rSearch = 2,
labelList = seglist, rad = rep(2, length(dim(ref))), verbose = TRUE
)
plot(ref, pp$segmentation)
plot(pp$intensity)
} # }