The categorical structural similarity index measure for 2D or 3D categorical or
binary images for multiple scales. The default is to compute over 5 scales.
This determines whether this is a 2D or 3D image and applies the appropriate
windowing, weighting, and scaling. Additional arguments can be passed.
This is a wrapper function for the 2D and 3D functions whose functionality
can be accessed through the ... arguments. This function is a wrapper for the
catmssim_2d(), catmssim_3d_slice(), and
catmssim_3d_cube()
functions.
catsim(
x,
y,
...,
cube = TRUE,
levels = NULL,
weights = NULL,
method = "Cohen",
window = NULL
)a binary or categorical image
additional arguments, such as window, can be passed as well as arguments for internal functions.
for the 3D method, whether to use the true 3D method
(cube or catmssim_3d_cube())
or compute the metric using 2D slices which are then averaged
(catmssim_3d_slice()). By default, TRUE,
which evaluates as a cube. FALSE will treat it as
2D slices.
how many levels of downsampling to use. By default, 5. If
weights is specified and this is left blank, the argument
will be inferred from the number of weights specified.
a vector of weights for the different scales. By default,
equal to rep(1,levels)/levels. If specified, there must
at least as many weights as there are levels and the first
levels weights will be used.
whether to use Cohen's kappa (Cohen),
Jaccard Index (Jaccard), Dice index (Dice),
accuracy (accuracy), Rand index (Rand),
Adjusted Rand Index (AdjRand or ARI), normalized mutual
information (NMI or MI), or adjusted mutual information
(AMI) as the similarity index.
Note Jaccard and Dice should only be used on binary data.
by default 11 for 2D and 5 for 3D images, but can be
specified as a vector if the window sizes differ by dimension.
The vector must have the same number of
dimensions as the inputted x and y.
a value less than 1 indicating the similarity between the images.
set.seed(20181207)
dim <- 16
x <- array(sample(0:4, dim^3, replace = TRUE), dim = c(dim, dim, dim))
y <- x
for (j in 1:dim) {
for (i in 1:dim) y[i, i, j] <- 0
for (i in 1:(dim - 1)) y[i, i + 1, j] <- 0
}
catsim(x, y, weights = c(.75, .25))
#> [1] 0.791647
# Now using a different similarity score
catsim(x, y, levels = 2, method = "accuracy")
#> [1] 0.7449315
# with the slice method:
catsim(x, y, weights = c(.75, .25), cube = FALSE, window = 8)
#> [1] 0.7858388