Detecting planes in binary images

Algorithms
1

I’m trying to detect certain type of artifacts in binary volumes I’m working with, namely step-like transitions between 0 and 1, but only in the three orthogonal planes XY/XZ/YZ. Example:

Screenshot from 2020-05-25 09-01-14

The surface is generally rough and doesn’t end sharply, but in this case it’s truncated in Z direction (the volume is rotated in the screenshot).

What ITK filters can I use to do this efficiently? I don’t need to segment those edges, I just need to know whether they exist in the volume. I would also need to somehow control the size threshold: it’s fine if a 2x2 voxel area vanishes abruplty, but a 20x20 step is something I want to know about.

A simple idea I tried is to convolve an edge-detecting kernel and threshold the output’s absolute intensity (this one is for detecting edges in Z direction only):

    kernel_np = np.zeros((2, 20, 20), dtype=np.int16)
    kernel_np[0, :, :] = -1
    kernel_np[1, :, :] = 1
    kernel_sitk = sitk.GetImageFromArray(kernel_np)

    output = sitk.Convolution(input, kernel_sitk)
    output = sitk.Abs(output)
    output = sitk.IntensityWindowing(output, 200.0)

I think I can get it to work with some fine tuning, but it’s quite slow. The binary volumes I’m analysing are of the order of 500x500x500.

2

That is a huge kernel for convolution. Try a smaller one.

1 Like
3

Good idea. I was thinking that I could filter out smaller flat areas by making the kernel big, but that’s obviously inefficient.

Can I use the fact that the volumes I’m working with are sparse (>90% zeros) or that they’re all binary?

4

Most label volumes are sparse, and many are binary. Not easy to take advantage of that in this case. Maybe check whether any of the morphological filters help you somehow?

1 Like