ANTsRNpy

Write a single or multi-channel antsImage to a numpy vector or matrix. Read a numpy vector or matrix back into a single or multi-channel antsImage.

Package dependencies: ANTSR, RcppCNPy

Installation:

devtools::install_github("stnava/ANTsRNpy")

See the help for the two functions:

readNumpyToANTsImage
writeANTsImageToNumpy

The help has examples for both python and R. Here are brief examples:

First, R (single-channel example):

library( ANTsRNpy )
ofn = "/tmp/temp.npy"
img = ANTsR::makeImage( c( 5, 6, 7 ), rnorm( 5*6*7) )
writeANTsImageToNumpy( img, ofn )
img3 = readNumpyToANTsImage( ofn, img )
as.array( img )[ 5, 2, 4 ]

Now python (corresponding to single-channel example from above):

# to read in python, do
from PIL import Image
import numpy as np
from scipy.misc import toimage
ofn = "/tmp/temp.npy"
data = np.load( ofn )
array = np.reshape( data, [ 7, 6, 5 ] )
# then we have
array[ 3, 1, 4 ] # index in python
# as.array( img )[ 5, 2, 4 ] # index in R
# so the python indices are in reverse order and minus one compared to R
toimage(array[:,:,2]).show()

So, from above, we can see that the necessary index reordering for 2D and 3D is:

2D: if R is m * n then the numpy data in python is n * m
3D: if R is m * n * o then the numpy data in python is o * n * m

The examples illustrate this and the differences induced by the 1-based indexing in R vs. 0-based indexing in python.

why do we need this tool? we want to create multichannel image ground truth data in R and read into python. however, R currently lacks the ability to write numpy data with dimensionality greater than 2 (i.e. matrices are ok but multi-dimensional arrays cannot be written). this is due to limits in the current implementation of RcppCNPy which may, in the future, be overcome. In the meantime, the current approach allows us to write either single or multi channel images into numpy vectors and matrices. we then aggregate and pickle the numpy data in python to create sharable training and testing datasets.