Misaligned intensity volume and ground truth mask

Vojtech_Benda · October 2, 2024, 4:28pm

I’m trying to extract volume sizes of a liver and tumors from their ground truth volume masks. I’m using the LabelIntensityStatisticsImageFilter, but for one of the inputs it throws me an error:

RuntimeError: Exception thrown in SimpleITK LabelIntensityStatisticsImageFilter_Execute: /tmp/SimpleITK-build/ITK-prefix/include/ITK-5.4/itkImageToImageFilter.hxx:215:
ITK ERROR: LabelImageToStatisticsLabelMapFilter(0x5632f36abe80): Inputs do not occupy the sam	e physical space! 
	InputImage Origin: [-1.0000000e+00, -1.0000000e+00, 1.0000000e+00], InputImage_1 Origin: [-2.4910001e+02, 2.4902318e+02, -6.5100000e+02]
		Tolerance: 1.0000000e-06
	InputImage Spacing: [1.0000000e+00, 1.0000000e+00, 1.0000000e+00], InputImage_1 Spacing: [9.7656202e-01, 9.7656202e-01, 2.5000000e+00]
		Tolerance: 1.0000000e-06
	InputImage Direction: -1.0000000e+00 0.0000000e+00 0.0000000e+00
	0.0000000e+00 -1.0000000e+00 0.0000000e+00
	0.0000000e+00 0.0000000e+00 1.0000000e+00
	, InputImage_1 Direction: 1.0000000e+00 0.0000000e+00 0.0000000e+00
	0.0000000e+00 -1.0000000e+00 0.0000000e+00
0.0000000e+00 0.0000000e+00 1.0000000e+00

	Tolerance: 1.0000000e-06

Viewing the inputs confirms that they do not at all occupy the same physical space (see image below), caused by non matching direction, spacing and origin.
wrong_space

I assumed simply resampling the ground truth volume using the intensity volume as reference image would align them (example code below), but that didn’t work.

vol = sitk.ReadImage(vol_path)
gt = sitk.ReadImage(gt_path)
gt_res = sitk.Resample(gt, vol, interpolator=sitk.sitkNearestNeighbor, outputPixelType=gt.GetPixelID())

The resampled volume’s origin, spacing and direction is correct, but it is empty - all voxels have value 0. The same result happens when using any of Slicer’s resample filters. Am I missing something or is this approach straight up wrong?

dchen · October 2, 2024, 5:44pm

No, resampling isn’t going to align them. Is a matter of the label image losing the origin, spacing and direction from original image? If so, you could copy them over using the Get/Set methods for Origin, Direction and Spacing.

zivy · October 2, 2024, 6:30pm

Hello @Vojtech_Benda,

Per @dchen’s advice, copy the information from the volume to the segmentation:

gt.CopyInformation(vol)

Vojtech_Benda · October 2, 2024, 6:41pm

Many thanks @dchen and @zivy. That fixed the misalignment. It also fixed the squished appearance of the ground truth mask.

Didn’t expect it to be that easy. Thanks a lot