ENH: Apply the ITK history of CannyEdgeDetectionImageFilter to its recursive Gaussian copy

[hjmjohnson]

CannyEdgeDetectionRecursiveGaussianImageFilter is a 2009 copy of itk::CannyEdgeDetectionImageFilter that never received ITK's later changes. This PR applies them: 5 bug-fix commits that keep their original ITK authors, then 7 themed commits (API, threading, performance, modernization, style, docs, tests) that credit the original authors with Co-Authored-By, and a final commit that trims the remaining differences. The two files now differ only in the smoothing filter. Part of #63.

For reviewers:

• Fixes: edges no longer depend on the start index of the image region (ITK bug 8729), repeated and successive Update() calls are correct, and a downstream consumer works when the release-data flag is on (the internal pipeline is isolated from the external one). Output is bit-identical to the current code on 2,100 random 2D and 3D images when none of those cases apply.
• API removed: SetThreshold/GetThreshold and SetOutsideValue/GetOutsideValue, as ITK did in 2011. OutsideValue was stored but never used. The callers in this repository (three feature generators and one example) are updated. A search of the ITK forest testbed repositories and Slicer found no other users of the class.
• Compare with ITK: side-by-side diff against ITK main of the two filter files (red is this repository, green is itk::CannyEdgeDetectionImageFilter at ITK 50c5b33430e with the class name replaced and this repository's .clang-format applied). The branch itk-main-reference-canny exists only for this comparison and is not part of the PR.
• Every commit builds and passes the full test suite on ITK 6 and ITK 5.4.7, so git bisect works. Any merge method is fine; "Squash and merge" would drop the per-commit authorship.

Test results

Environment	Result
macOS arm64, ITK 6 (5d31af3cfdf)	all 13 commits build; 43/43 tests pass at commits 1 to 11, 51/51 at commits 12 and 13 (the new tests arrive in commit 12); rebuilt from scratch at the head, 51/51
macOS arm64, ITK 5.4.7 (release-5.4 build)	the same: all 13 commits build; 43/43 and then 51/51; rebuilt from scratch at the head, 51/51
Linux x86_64, clang 18, ITK 6, no sanitizer	at the head: Canny tests 1, 5, 6, 7, and 8 pass
Linux x86_64, clang 18, ITK v6.0b02 (the ITK 6 of the Python wheel jobs)	full suite 51/51 at the head; the previous head, which guarded on ITK_VERSION_MAJOR, failed to compile there
Linux x86_64, clang 18, ITK 6, ASan and UBSan	at the head: full suite 50/51, Canny tests 1 to 8 pass; the one failure is itkLandmarksReaderTest1, which never frees the list returned by GetChildren() and fails the same way on main (28f2cf5)
Linux x86_64, clang 18, ITK 6, ThreadSanitizer	at the head: tests 1, 5, 6, 7, and 8 and the 2,100-case harness: 0 warnings
Linux x86_64, valgrind memcheck	tests 1, 5, 6, 7, and 8: 0 errors, 0 definite or indirect leaks
GitHub Actions (ITK 5.4.7 and ITK 6 on Windows/macOS/Linux, plus Python wheels)	at the head 31a2d1c: 27 checks pass, 2 skipped (the notebook jobs, also skipped on main), 0 failed; the two ARM ITK 5.4.7 Python jobs hit a runner DNS error (Could not resolve host: github.com) and passed on rerun

The 500-run ASan loop and the valgrind runs were made on the tree at commit 12; everything else in the table was rerun on the head. Commit 13 changes no behavior (an unused member removed, one getter made non-const, GaussianImageFilterType made protected as in ITK, comments and tags, and the feature-guarded PrintSelf and iterator declarations), and the 2,100-case harness hashes are identical to main at the head on ITK 6 and ITK 5.4.7 on macOS and on ITK 6 on Linux, and the full suite was rerun on it with ITK 6 and ITK 5.4.7, which compile the two branches of each version guard respectively.

Two feature-generator tests (itkCannyEdgesDistanceFeatureGeneratorTest1 and itkCannyEdgesDistanceAdvectionFieldFeatureGeneratorTest1) report a UBSan runtime error when it is made fatal (halt_on_error=1). The report is inside ITK's BinaryContourImageFilter (ScanlineFilterCommon::SetupLineOffsets, a null-pointer offset in ConstNeighborhoodIterator), not in this filter. Both tests pass otherwise.

The eight new itkCannyEdgeDetectionRecursiveGaussianImageFilterTest* tests mirror ITK's itkCannyEdgeDetectionImageFilterTest and Test2: a baseline comparison and an A/B re-execution check on cthead1.png (new 197 KB fixture), the same A/B check on the 3D lesion volume, and tests of region start index, successive inputs of different sizes, parameters, hysteresis linking, error handling, and a downstream pipeline with the release-data flag on (Test8, which throws a buffered-region exception on main and passes here). The baseline PNG was generated by the filter before the ITK history was applied.

Equivalence with the current code

A harness runs the filter on 2,100 random images: 300 2D images (4 to 73 pixels per side) and 120 3D images (4 to 31 pixels per side), each at 1, 2, 3, 7, and 16 work units, with random sigma and thresholds. 620 of the cases contain edges. The FNV-1a hash of the output buffer is identical between the code on main and this branch:

• on ITK 6 (5d31af3cfdf), 2,100 of 2,100 cases;
• on ITK 5.4.7, 2,100 of 2,100 cases.

The output does not depend on the number of work units: 420 image and seed groups, 0 mismatches.

The harness excludes the four situations this PR fixes (non-zero start index, repeated Update(), successive inputs, pipelines with the release-data flag), where the old code was wrong. Those are covered by Test2, Test5, Test7, and Test8, which fail on main and pass here.

The commits

#	Commit	Original authors
1	BUG non-zero start index (InBounds replaced by IsInside), ITK bug 8729	Bradley Lowekamp
2	BUG integer types on 64-bit Windows, ITK bug 9260	Luis Ibanez
3	BUG correct results when called multiple times (stale edge-linking node), ITK bug 9431	Bradley Lowekamp
4	BUG ITK_TEMPLATE_EXPORT on ListNode	Mayeul Chassagnard
5	BUG isolate the internal filter from the external pipeline	Bradley Lowekamp
6	ENH API cleanups: remove legacy code and Threshold/OutsideValue, const getters, size types	7 ITK commits
7	ENH ITK 5 threading: MultiThreaderBase, ParallelizeImageRegion, work units, ProgressTransformer	7 ITK commits
8	PERF hoist accessor calls out of inner loops	3 ITK commits
9	COMP C++ modernization	44 ITK commits
10	STYLE formatting, naming, and include conventions	17 ITK commits
11	DOC comments, license header, and documentation	18 ITK commits
12	ENH the eight new tests, fixture, and baseline	this PR
13	ENH minimize the remaining differences from itk::CannyEdgeDetectionImageFilter	this PR

Commits 1 to 5 are the bug fixes. Their dependencies are small, so they are applied first: commit 4 adds only the ListNode half of ITK's 2016 fix (the filter class already had the keyword), and commit 5 includes the private output pointer from ITK's 2012 performance commit that the isolation fix builds on.

Commits 6 to 11 each list the ITK commits they contain (hash, date, and subject) in the message, and carry a Co-Authored-By trailer for every original ITK author. Together with the bug-fix commits, 101 ITK commits are listed.

Two ITK changes are handled differently, on purpose:

• ITK's 2026 PrintNumericTrait adoption (2026-05-08) is applied only when the ITK being built against has it. ITK_VERSION_MAJOR cannot tell, because v6.0b02 and current main both report 6.0.0, so CMakeLists.txt compares ITK_VERSION_TWEAK (ITK's commit date) with the date the feature landed and defines LST_ITK_HAS_PRINT_NUMERIC_TRAIT. Without the macro, including for downstream code that does not use this CMake file, PrintSelf keeps the static_cast form (commit 13).
• ITK's iterator class-template-argument-deduction commits (2026-02-27) do not compile on ITK 5.4 or v6.0b02, so commit 9 spells out ImageRegionIterator<TOutputImage> at four sites, and commit 13 restores ITK's ImageRegionIterator spelling behind LST_ITK_HAS_ITERATOR_CTAD, defined the same way from ITK_VERSION_TWEAK.

How the history was produced

The first LST version equals ITK's revisions from late 2008 once the class name is normalized: the header from bd48a2188a8 and the implementation from 5f06f26925b, with 0 differing lines. For each of the 125 later ITK commits that touch the header or implementation, the before and after versions were renamed, formatted with this repository's .clang-format, and merged into the LST file with a three-way merge, replaying ITK's history in order. 22 commits produced no change because LST had already made the same edit. Conflicts were resolved by keeping LST's smoothing code (Sigma, SmoothingRecursiveGaussianImageFilter) and taking ITK's change elsewhere, with a brace-depth check after every commit.

The five bug fixes were then re-created as the first five commits. Commits 6 to 11 group the remaining ITK commits by theme: the intermediate trees are built by replaying each cumulative set of groups from commit 5 (resolving conflicts toward the commit's side), then repairing the few spots that did not compile. The tree at commit 12 equals the full, in-order replay byte for byte; commit 13 then removes the last avoidable differences. Every intermediate commit was built and tested as described above.

Distance to ITK's current files, after normalizing the class name and formatting (differing lines):

File	Before	After
itkCannyEdgeDetectionRecursiveGaussianImageFilter.h	265	65
itkCannyEdgeDetectionRecursiveGaussianImageFilter.hxx	528	83

What remains, including the lines of the feature guards, is the Sigma versus Variance and MaximumError smoothing parameters, the Sigma accessors, and doc comments. To compare:

N=s/CannyEdgeDetectionRecursiveGaussianImageFilter/CannyEdgeDetectionImageFilter/g
vimdiff <(sed "$N" include/itkCannyEdgeDetectionRecursiveGaussianImageFilter.hxx) \
        $ITK/Modules/Filtering/ImageFeature/include/itkCannyEdgeDetectionImageFilter.hxx

Tooling and AI assistance

The replay was scripted and the conflicts were resolved with assistance from Claude Code (Anthropic). The author reviewed the results and is responsible for them. The scripts, harness, and result files are kept by the author and can be shared on request.

[hjmjohnson]

@dzenanz I was trying to preserve maximal attribution and credit for all the changes. But I now see that is not all that useful, and I will squash this into a few commits. These files are copy-and-paste files with algorithm components (smoothing components swapped) from 2009, so conceptually it's really not that much of a change, but it occurred over 17 years.

[hjmjohnson]

@dzenanz The main goal of this was a learning experience for me trying some new approaches out. The goal was to minimize the difference between the local LSTK itkCannyEdgeDetectionRecursiveGaussianImageFilter.h and the original copied itk itkCannyEdgeDetectionImageFilter.h.

The ITK itkCannyEdgeDetectionImageFilter.h has over 100 code updates, bug fixes, and style improvements; I was trying to preserve attribution for those who contributed to the fixes.

In the end, many of the removals of code are provably no-ops that are copy-and-paste cruft, or building out of tests from ITK/CannyEdgeDetectionImageFilter to add coverage for the LSTK/CannyEdgeDetectionRecursiveGaussianImageFilter.