Use simple ITK labelShapeStatisticsFilter for computation of Shape features

[aim-harvard]

Use simple ITK labelShapeStatisticsFilter for computation of Shape features

[JoostJM]

After a quick view and test today, I think the new features are in labelshape statistics. Also max3Ddiameter and volume are in there. The rest of the features (with the exception of surface area) are simple calculations that are based on volume and surface area.
Just one problem so far, I can't find simple ITK's equivalent of surface area
there is a function that calculates the perimeter, but this is a distance, not a surface. (And very different from the matlab code anyhow, by about 30%).

[JoostJM]

I have tested the use of LabelShapeStatisticsImageFilter.GetFeretDiameter (which should be equal to max3Ddiameter), this worked. Additionally, I discovered the cause of the differences between matlab and pyradiomics, it is due to a bug in the pyradiomics code, after the fix, the pyradiomics code was also computing the correct distance.

[JoostJM]

The differences between surface area calculated in matlab and pyradiomics are very large (+/- 500), so there is a bug somewhere probably.
swapping around pixelspacing does not have a large effect, so not solely responsible for large differences.

[fedorov]

Thanks for checking! Let's keep thinking about this.

[JoostJM]

@fedorov currently working on a slightly different implementation to count voxel faces, will update...

[JoostJM]

@fedorov good news, my revision drastically reduced the size of the error. still 15-30% though

I think the differences are now more likely to be explained as due to the different type of defining the surface...

[fedorov]

Can you make a synthetic example where we know the actual surface area, and add this as a regression test?

[JoostJM]

@fedorov I'll try, should be possible. How about a sphere, radius 10mm, in sitk image of [1,1,1] spacing?

[fedorov]

I would start with a cube, and test with isotropic spacing, then with different slice thickness, then if those tests pass, maybe make a star-like shape, or something with slowly increasing level of complexity, but still manageable to calculate true answer manually.

I have to admit, I can't calculate voxelized surface area of a sphere in my head! :-(

[JoostJM]

Good idea!

[JoostJM]

Resolved by #138