Slicing a 3D array vs projecting it

Question

Link to array file

I am new to image processing and trying to understand projection. So when we have a 3D image, it is simply a 3D numpy array and i view it by slicing a 2D array out of the 3D array. In order to do a orthogonal projection, i simply sum up the arrays along one axis. This is my code for doing that:

import numpy as np
import matplotlib.pyplot as plt

#shape of the 'image' array is (256, 256, 176) so i am assuming there are 256 slices and row and column of each slice is 256x176
image = np.load('brain_ct.npy')

#when i sum along axis 0 or axis 1, the image shows up as blank with some warnings
#(Warning: converting a masked element to nan)
#only summing up along axis 2 works (not sure why is that)
collapsed = np.sum(image, axis=2)

plt.imshow(collapsed, cmap='gray')

I am not sure why summing up along axis 0 and 1 does not work. Also, i am not sure what view (top/down/side) axis 2 is giving me? Finally, if summing up along one axis gives orthogonal projection, how can i go about doing an oblique and perspective projection. Is there any transformation matrix that i have to multiply the image coordinates with?

Thanks everyone.

Answer 1

The error appears to show that you are loading a numpy.ma.MaskedArray rather than a NumPy array. However, I don't understand why you would not get the error along axis=2, since any masked points will be projected along each dimension. It could be that you have a whole plane masked, and then summing along 0 will just give you an array of nans. I am not sure but it could be that np.sum does not handle masked arrays properly, and you should instead use image.sum() , which might know better how to handle the masks. If you share your .npy file we can provide a more personalised response.

In general, summing is one way to project, but you could also do max() , which is the most common form of volumetric projection.

Regarding oblique projections, that's far more than can be succinctly summarised in a SO answer, but in general yes, you want to transform your image and then do the projection. You can look at the scipy.ndimage module for more information on this.