DICOM importer does not check for missing slices / inconsistent slice spacing

[jmsmkn]

See comic/grand-challenge.org#1301

[nlessmann]

This would be a very useful feature since inconsistent slice spacing is quite prevalent in some of our datasets.

The DICOM standard supports this but the formats that we actually work with do not - so in my opinion, the solution should be to use all of the present data to interpolate the data onto a regular grid with the most common spacing.

[pkcakeout]

The DICOM standard supports this but the formats that we actually work with do not

Spoken like a true CT-centrist ;-) We also support MR on grand-challenge now and inconsistent slice spacings are way more common for MR than for CT if I am not mistaken. So it still needs to be considered, as I wrote in my in the original issue.

It is also fine to reject those, but we then need to specify what is actually supported.

[cjacobs1]

I would argue that an import error should be produced, reporting back that no single volumetric dataset could be created from the set of DICOM files. I think we should not interpolate data by default.

[cjacobs1]

I have prepared some test data for this. It can be found here: \\umcsanfsclp01.umcn.nl\radng_diag\rse\projects\rse-panimg\testdata

[jmsmkn]

The spacing is determined here, you should be able to raise any exception and it be passed along and the study will be rejected:

rse-panimg/panimg/image_builders/dicom.py

Lines 173 to 189 in 9e46f1f

	# Multiple slices, average spacing between slices
	avg_origin_diff = origin_diff / n_diffs
	z_i = np.linalg.norm(avg_origin_diff)
	# Use orientation of the coordinate system to determine in which
	# direction the origins of the individual slices should move.
	# Use the dot product to find the angle between that direction
	# and the spacing vector - this tells us whether the order of
	# the slices is correct or should be reversed
	z_direction = self.direction @ np.ones(self.dimensions)
	if self.dimensions > 3:
	avg_origin_diff = np.pad(
	avg_origin_diff,
	pad_width=(0, self.dimensions - avg_origin_diff.size),
	mode="constant",
	)
	z_order = np.sign(np.dot(avg_origin_diff, z_direction))

[cjacobs1]

Thanks, I'll work on a fix and prepare a PR.