ENH: Ingest AnisotropicDiffusionLBR into Modules/Filtering

[hjmjohnson]

Moves AnisotropicDiffusionLBR from a configure-time remote fetch into the ITK source tree at Modules/Filtering/AnisotropicDiffusionLBR/. Full upstream git history and blame preserved. Relates to #6060.

Approach

Per reviewer consensus on #6061 and #6085: remote modules go directly into their appropriate group — no intermediate staging directory.

The upstream repo (ITKAnisotropicDiffusionLBR) is grafted via git merge --allow-unrelated-histories --no-ff of a filter-repo'd clone that rewrites paths under Modules/Filtering/AnisotropicDiffusionLBR/. The merge commit is byte-identical to the upstream tree (just path-prefixed); all artifact removal happens in a separate STYLE commit.

Commit structure (5 commits)

1. COMP: Bump AnisotropicDiffusionLBR to upstream/main tip — updates GIT_TAG to upstream HEAD (203260b9)
2. ENH: Ingest ITKAnisotropicDiffusionLBR into Modules/Filtering — merge commit, byte-identical to upstream tree
3. STYLE: Remove non-ITK artifacts from ingested AnisotropicDiffusionLBR — removes .github/, pyproject.toml, CTestConfig.cmake, .clang-format, .gitattributes, LICENSE (Apache-2.0 duplicate)
4. COMP: Remove AnisotropicDiffusionLBR.remote.cmake; now in-tree — deletes the remote fetch declaration
5. COMP: Fix pre-commit hook failures for AnisotropicDiffusionLBR — gersemi, trailing whitespace, EOF fixes

Verification

• 136 upstream commits preserved across the merge boundary
• 12 distinct authors surface in git log (Matt McCormick, Bradley Lowekamp, Dženan Zukić, Hans Johnson, Jon Haitz Legarreta, Tom Birdsong, etc.)
• git blame walks through the merge boundary to original authors (2015+)
• Configure: clean
• Build: no errors, no warnings
• Tests: 6/6 pass
• Pre-commit: all hooks pass

AI disclosure

Ingestion tooling (ingest-remote-module.sh) and commit structure developed with Claude Code assistance. All commits reviewed and verified locally.

[hjmjohnson]

@thewtex — this PR demonstrates the process for ingesting remote modules directly into their appropriate group, per your feedback on #6061 and #6085. One PR per module, full upstream history preserved via git merge --allow-unrelated-histories. Module stays EXCLUDE_FROM_DEFAULT; CI opts in explicitly via configure-ci.

If this looks right, I'll proceed with the remaining ~30 Tier-A modules using the same pattern.

Overall ingestion plan

Approach (per reviewer consensus on #6061, #6085):

• Each remote module goes directly into its existing group (Modules/Filtering/, Modules/IO/, etc.)
• One PR per module
• Full upstream git history preserved — git blame walks across the merge boundary
• Upstream repo archived on GitHub after merge

Commit structure per module (7 commits):

1. COMP: Bump <Name> to upstream/<branch> tip — sync to latest upstream SHA
2. ENH: Ingest ITK<Name> into Modules/<Group> — merge commit, byte-identical to upstream tree (just path-prefixed)
3. STYLE: Remove non-ITK artifacts from ingested <Name> — .github/, pyproject.toml, CTestConfig.cmake, .clang-format, .gitattributes, redundant LICENSE
4. COMP: Remove <Name>.remote.cmake; now in-tree — delete the remote fetch declaration
5. COMP: Fix pre-commit hook failures for <Name> — gersemi, trailing whitespace, EOF newlines
6. STYLE: Add missing trailing newlines to .md5 data files — ghostflow compliance
7. ENH: Enable <Name> in CI via configure-ci — add -DModule_<Name>:BOOL=ON to pixi configure-ci

What stays independent (not ingested):
RTK, CudaCommon, Cleaver, SCIFIO, Shape, TractographyTRX, WebAssemblyInterface, SphinxExamples, TubeTK, Ultrasound, and modules with external library deps (TotalVariation, IOOMEZarrNGFF, IOOpenSlide, VkFFTBackend, MeshToPolyData, IOTransformDCMTK).

Proposed wave order:

Wave	Modules	Group
1	AnisotropicDiffusionLBR, FastBilateral, LabelErodeDilate, GenericLabelInterpolator, SplitComponents, PolarTransform, MultipleImageIterator, HigherOrderAccurateGradient, ParabolicMorphology, MorphologicalContourInterpolation, SmoothingRecursiveYvvGaussianFilter, Cuberille, MeshNoise, SubdivisionQuadEdgeMeshFilter	Filtering
1	IOMeshSTL, IOMeshMZ3, IOFDF	IO
2	BoneEnhancement, BoneMorphometry, TextureFeatures, IsotropicWavelets, Montage, Thickness3D, Strain, PhaseSymmetry, SimpleITKFilters	Filtering
2	IOScanco, MGHIO	IO
2	GrowCut	Segmentation
2	RANSAC, VariationalRegistration	Registration

Ingestion script (ingest-remote-module.sh)

# Usage:
./ingest-remote-module.sh <ModuleName> <DestGroup> [OPTIONS]

# Examples:
./ingest-remote-module.sh AnisotropicDiffusionLBR Filtering
./ingest-remote-module.sh IOMeshSTL IO
./ingest-remote-module.sh GrowCut Segmentation --dry-run

# Options:
#   --dry-run          Preview without executing
#   --skip-bump        Don't update GIT_TAG to upstream tip
#   --skip-precommit   Don't run pre-commit fixes
#   --upstream-url     Override GIT_REPOSITORY from .remote.cmake
#   --upstream-branch  Override upstream branch (default: auto-detect)
#   --itk-source       Override ITK source root detection

What it automates:

1. Parses Modules/Remote/<Name>.remote.cmake for upstream URL and SHA
2. Bumps GIT_TAG to the current upstream tip (if out of date)
3. Clones upstream, runs git filter-repo --to-subdirectory-filter to rewrite paths
4. Merges into ITK with --allow-unrelated-histories --no-ff (byte-identical to upstream)
5. Removes non-ITK artifacts (CI configs, packaging scaffolding) in a separate STYLE commit
6. Deletes the .remote.cmake file
7. Runs pre-commit hooks to convergence
8. Adds -DModule_<Name>:BOOL=ON to configure-ci in pyproject.toml

Prerequisites: git-filter-repo (pip install git-filter-repo), pixi, clean working tree.

What it does NOT do: push, open PRs, or modify source code. All improvements (clang-tidy, CMake modernization, test renames) are deferred to follow-up PRs.

[greptile-apps]

Greptile Summary

This PR ingests AnisotropicDiffusionLBR from its remote fetch location into Modules/Filtering/AnisotropicDiffusionLBR/, preserving the full upstream git history via a filter-repo-rewritten merge commit. The remote .cmake file is deleted, pyproject.toml is updated to enable the module in CI, and upstream-only artifacts (.github/, pyproject.toml, CI configs) are stripped in a separate STYLE commit.

• P1: itkStructureTensorImageFilter.h has using Superclass = ImageToImageFilter<TImage, TImage> but the class inherits from ImageToImageFilter<TImage, TTensorImage> — these are distinct template instantiations and the alias misrepresents the base class.

Confidence Score: 4/5

Safe to merge after addressing the Superclass typedef bug in itkStructureTensorImageFilter.h

One P1 finding: the Superclass alias in itkStructureTensorImageFilter resolves to the wrong template instantiation (TImage vs TTensorImage), which is a latent type-correctness defect even though the current build passes. The P2 on itk-module.cmake I/O deps is cosmetic. The ingestion mechanics, history preservation, and CI configuration are all correct.

Modules/Filtering/AnisotropicDiffusionLBR/include/itkStructureTensorImageFilter.h (wrong Superclass alias)

Important Files Changed

Filename	Overview
Modules/Filtering/AnisotropicDiffusionLBR/include/itkStructureTensorImageFilter.h	Incorrect Superclass typedef — aliases ImageToImageFilter<TImage, TImage> instead of the actual base ImageToImageFilter<TImage, TTensorImage>
Modules/Filtering/AnisotropicDiffusionLBR/itk-module.cmake	ITKIOSpatialObjects and ITKMetaIO listed in module DEPENDS but no public headers require them; likely should be in TEST_DEPENDS or removed
Modules/Filtering/AnisotropicDiffusionLBR/CMakeLists.txt	Standard in-tree module build file; version hardcoded to 5.1.0 with a comment noting it should track ITK
Modules/Remote/AnisotropicDiffusionLBR.remote.cmake	Remote fetch declaration correctly removed now that the module is in-tree
Modules/Filtering/AnisotropicDiffusionLBR/include/itkAnisotropicDiffusionLBRImageFilter.h	Base non-linear anisotropic diffusion filter; clean ITK-style header with correct type aliases
Modules/Filtering/AnisotropicDiffusionLBR/include/itkCoherenceEnhancingDiffusionImageFilter.h	CED/EED subclass; clean header with properly overridden EigenValuesTransform
Modules/Filtering/AnisotropicDiffusionLBR/include/itkLinearAnisotropicDiffusionLBRImageFilter.h	Linear diffusion core filter with LBR stencil arithmetic; correct ITK conventions
Modules/Filtering/AnisotropicDiffusionLBR/test/CMakeLists.txt	15 test cases across 2D/3D images with proper DATA{} references and baseline comparisons
pyproject.toml	Adds -DModule_AnisotropicDiffusionLBR:BOOL=ON to CI configure, correctly enabling the now in-tree module
Modules/Filtering/AnisotropicDiffusionLBR/wrapping/CMakeLists.txt	Wrapping configured with WRAPPING_SUBMODULE_ORDER ensuring base class wraps before subclass; auto-discovers itkCoherenceEnhancingDiffusionImageFilter.wrap

Class Diagram

%%{init: {'theme': 'neutral'}}%%
classDiagram
    class ImageToImageFilter~TImage,TOutput~ {
        +GenerateData()
    }
    class LinearAnisotropicDiffusionLBRImageFilter~TImage,TScalar~ {
        +SetInputImage()
        +SetInputTensor()
        +SetMaxDiffusionTime()
        +GenerateData()
    }
    class AnisotropicDiffusionLBRImageFilter~TImage,TScalar~ {
        +EigenValuesTransform()
        +ComputeDiffusionTensors()
        +GenerateData()
    }
    class CoherenceEnhancingDiffusionImageFilter~TImage,TScalar~ {
        +EigenValuesTransform()
        -g_CED()
        -g_EED()
    }
    class StructureTensorImageFilter~TImage,TTensorImage~ {
        +GenerateData()
    }

    ImageToImageFilter <|-- LinearAnisotropicDiffusionLBRImageFilter
    ImageToImageFilter <|-- AnisotropicDiffusionLBRImageFilter
    ImageToImageFilter <|-- StructureTensorImageFilter
    AnisotropicDiffusionLBRImageFilter <|-- CoherenceEnhancingDiffusionImageFilter
    AnisotropicDiffusionLBRImageFilter o-- LinearAnisotropicDiffusionLBRImageFilter : uses
    AnisotropicDiffusionLBRImageFilter o-- StructureTensorImageFilter : uses

Loading

_{Reviews (1): Last reviewed commit: "ENH: Enable AnisotropicDiffusionLBR in C..." | Re-trigger Greptile}

[greptile-apps]

Incorrect Superclass type alias

Superclass is aliased to ImageToImageFilter<TImage, TImage>, but the class actually inherits from ImageToImageFilter<TImage, TTensorImage>. These are two different template instantiations: TImage ≠ TTensorImage (the default TTensorImage is Image<SymmetricSecondRankTensor<...>, N>, not TImage). Any downstream code that resolves StructureTensorImageFilter::Superclass::OutputImageType will get the wrong type.

Suggested change

	using Superclass = ImageToImageFilter<TImage, TImage>;
	using Superclass = ImageToImageFilter<TImage, TTensorImage>;

[greptile-apps]

Questionable module-level I/O dependencies

ITKIOSpatialObjects and ITKMetaIO appear in module DEPENDS, but none of the public headers include anything from these modules (no SpatialObject or MetaIO headers are pulled in). I/O used only in tests should live under TEST_DEPENDS, not DEPENDS, to avoid forcing downstream consumers to link against unused I/O libraries. If these are genuinely unused, they can be removed outright; if they are needed for test I/O (e.g. .vtk files), ITKIOVTK would be the correct dependency and it should go in TEST_DEPENDS.

[blowekamp]

I don't think merging the full history is a good idea to all remote repositories. The remote repos have not had the same rigor as the main repo, larger file ( such as those used for papers may have been included, or perhaps some copyright material). There could also have been a long initial development period too. If this is done with most remotes then I think the ITK git history may become to bloated.

There may be some git history filter that could be run to reduce it to the files that are important ( main cxx,h file ) but I am unsure.

Also it appear the remote CMake code is still here.

[hjmjohnson]

There may be some git history filter that could be run to reduce it to the files that are important ( main cxx,h file ) but I am unsure.

@blowekamp We could
-"squash" to 1 commit, (retain last-modified attribution, with reference to the original archived repo for deeper dives into attribution).

• or could have a history filter that removes large files,
• or removes all transient files that are not retained in the origin/main sandbox.
• or a combination of the above.

Could you help me define a set of characteristics that are most important to retain?

My desired characteristics:

1. retain last-modified attribution ... That's about it. The return on investment diminishes significantly after that. We have a complete history, it's just not as easy to find (but a vanishingly small number of people will need that).

[blowekamp]

For these particular file I see a comment "Created by Jean-Marie Mirebeau on 28/02/2014." but it does not look like he is a git author. I think this is a special case that may need to be created.

Perhaps we can just squash the commits and use primary author and co-author?

I think it will be important to remove extra files in this process. I don't think the paper and related material should be in ITK. This repo has an "Old" directory that I expect is not needed. There is an "example" directory with what appears to be a full content PNG file, which should be removed. I think an "example" directory with executables is reasonable on a per module basis, but I'm not aware of the current support in modules.