Add ImpactReg extension

[vboussot]

New extension

Tier 1

Any extension that is listed in the Extensions Catalog must fulfill these requirements.

• Extension has a reasonable name (not too general, not too narrow, suggests what the extension is for). The extension name should not start with Slicer (unless it explicitly provides a bridge between Slicer and a tool or library), because it would make it more difficult to find extensions if the name of many started with the same word.
• Repository name is Slicer+ExtensionName (except if the repository that hosts the extension can be also used without Slicer)
• Repository is associated with 3d-slicer-extension GitHub topic so that it is listed here. To edit topics, click the settings icon in the right side of "About" section header and enter 3d-slicer-extension in "Topics" and click "Save changes". To learn more about topics, read https://help.github.com/en/articles/about-topics
• Extension description summarizes in 1-2 sentences what the extension is usable (should be understandable for non-experts)
• Any known related patents must be mentioned in the extension description.
• LICENSE.txt is present in the repository root and the name of the license is mentioned in extension homepage. We suggest you use a permissive license that includes patent and contribution clauses. This will help protect developers and ensure the code remains freely available. MIT (https://choosealicense.com/licenses/mit/) or Apache (https://choosealicense.com/licenses/apache-2.0/) license is recommended. Read here to learn more about licenses. If source code license is more restrictive for users than MIT, BSD, Apache, or 3D Slicer license then describe the reason for the license choice and include the name of the used license in the extension description.
• Extension URL and revision (scmurl, scmrevision) is correct, consider using a branch name (main, release, ...) instead of a specific git hash to avoid re-submitting pull request whenever the extension is updated
• Extension icon URL is correct (do not use the icon's webpage but the raw data download URL that you get from the download button - it should look something like this: https://raw.githubusercontent.com/user/repo/main/SomeIcon.png)
• Screenshot URLs (screenshoturls) are correct, contains at least one
• Content of submitted json file is consistent with the top-level CMakeLists.txt file in the repository (dependencies, etc. are the same)
• Homepage URL points to valid webpage containing the following:
  • Extension name
  • Short description: 1-2 sentences, which summarizes what the extension is usable for
  • At least one nice, informative image, that illustrates what the extension can do. It may be a screenshot.
  • Description of contained modules: at one sentence for each module
  • Publication: link to publication and/or to PubMed reference (if available)
• Hide unused github features (such as Wiki, Projects, and Discussions, Releases, Packages) in the repository to reduce noise/irrelevant information:
  • Click Settings and in repository settings uncheck Wiki, Projects, and Discussions (if they are currently not used).
  • Click the settings icon next to About in the top-right corner of the repository main page and uncheck Releases and Packages (if they are currently not used)
• The extension is safe:
  • Does not include or download binaries from unreliable sources
  • Does not send any information anywhere without user consent (explicit opt-in is required)

Tier 3

Community-supported extensions.

• Documentation, tutorial, and test data are provided for most modules. A tutorial provides step-by-step description of at least the most typical use case, include a few screenshots. Any sample data sets that is used in tutorials must be registered with the Sample Data module to provide easy access to the user.
• Follows programming and user interface conventions of 3D Slicer (e.g., GUI and logic are separated, usage of popups is minimized, no unnecessary custom GUI styling, etc.)
• The extension can be successfully built and packaged on all supported platforms (Windows, macOS, Linux)
• Maintainers respond to issues and pull request submitted to the extension's repository.
• Maintainers respond to questions directly addressed to him/her via @mention on the Slicer Forum.
• Permissive license is used for the main functions of the extension (recommended Apache or MIT). The extension can provide additional functionality in optional components that are distributed with non-permissive license, but the user has to explicitly approve those before using them (e.g., a pop-up can be displayed that explains the licensing terms and the user has to acknowledge them to proceed).
• All requirements of tiers < 3.

Tier 5

Critically important extensions, supported by Slicer core developers. New Slicer Stable Release is released only if all Tier 5 extension packages are successfully created on all supported platforms.

• Slicer core developers accept the responsibility of fixing any issues caused by Slicer core changes; at least one Slicer core developer (anyone who has commit right to Slicer core) must be granted commit right to the extension's repository.
• Automated tests for all critical features.
• Maintainers respond to questions related to the extension on the Slicer Forum.
• All requirements of tiers < 5.

---

This PR adds IMPACT-Reg, a new extension enabling multimodal medical image registration directly inside 3D Slicer.

IMPACT-Reg integrates the IMPACT deep semantic similarity metric into the Elastix registration tools and exposes it through a user-friendly GUI.

---

Key functionality included

• Deep semantic registration (CT–CBCT, MR–CT, etc.)
• GPU-accelerated feature extraction when available
• Sequential execution of multiple presets
• Automatic warped volume and transform generation
• Built-in evaluation tools (landmarks, segmentations)
• Support for uncertainty analysis based on transform ensembles

---

The module depends on the KonfAI extension, which provides KonfAI App execution support.

[lassoan]

Thanks for your contribution! As I see, a custom build of Elastix binaries is downloaded from somewhere, but I could not easily find from where and how. Could you please clarify?

[lassoan]

FYI, on Windows I get this error when I try to run the registration:

Elastix execution failed.

Command:
C:\D\SlicerImpactReg\ImpactReg\Resources\bin\elastix-impact\elastix.exe -h
Return code: 3221225781

[vboussot]

The custom elastix binaries are simply a compilation of elastix with the IMPACT similarity metric enabled.

IMPACT relies on LibTorch at runtime to extract deep features.
As a result, the elastix build depends not only on the operating system, but also on the availability of a GPU driver.

For this reason, multiple pre-built elastix binaries are provided (CPU-only and CUDA-enabled).
When the elastix binaries are not yet installed (typically on the first use of SlicerImpactReg), the install.py script
(https://github.com/vboussot/SlicerImpactReg/blob/main/ImpactReg/Resources/bin/install.py
) is executed.

This script automatically selects the appropriate build, downloads it, and (if needed) fetches the matching LibTorch version and ensures it is correctly linked at runtime, so that elastix works out of the box.

[vboussot]

This return code usually indicates a missing runtime DLL.
Could you please provide the dependency listing for elastix.exe, for example using:

dumpbin /DEPENDENTS elastix.exe

This should help identify which runtime dependency is missing.

[lassoan]

On Windows (with Slicer-5.10), the custom-built Elastix packages are not self-contained - several pytorch DLLs are missing (even from the CPU package, such as fbgemm.dll). They work if I copy (and overwrite all existing DLLs) from "...\AppData\Local\slicer.org\3D Slicer 5.10.0\lib\Python\Lib\site-packages\torch\lib" to elastix-impact folder. Of course, the clean solution would be not to include torch binaries in the custom Elastix packages (not even for the CPU version, as it is incomplete anyway) and set the path before launching Elastix to make the loader find the torch DLLs (instead of copying them over).

FYI, on an RTX3060, registering MRBrainTumor1 to MRBrainTumor1 is really slow (it is at 35% after about 30 minutes). This is a problem, because Elastix basic bpsline takes <1min on CPU and it usually gives acceptable results.

[lassoan]

When I attempt to run elastix.exe from elastix-impact-windows-x86_64-shared-with-deps-cpu.zip, I get errors about missing fbgemm.dll and libiomp5md.dll.

[vboussot]

Thank you, indeed, the issue comes from the fact that the custom Elastix packages were not truly self-contained for the CPU version. In the initial packaging, I performed an overly aggressive pruning of PyTorch runtime libraries, assuming that some components (such as fbgemm.dll, libiomp5md.dll, etc.) were not required at runtime based on my local tests. Clearly, this assumption does not hold across all Windows environments.

I have since revised this approach:

No LibTorch libraries are shipped anymore with the custom Elastix binaries, for both CPU and CUDA builds.

The installation script now downloads the Impact release and the appropriate official LibTorch distribution (CPU or CUDA).

It then explicitly copies the required runtime libraries into a directory visible to the elastix binary.

Regarding the possibility of reusing PyTorch libraries installed through the PyTorch utils extension, unfortunately this installation is not under my control (exact version, CUDA runtime version). However, the Elastix/IMPACT binary requires a very specific LibTorch version (LibTorch 2.8.0, with or without CUDA 12.8).

As a result, while reusing torch/lib from a pip installation may avoid missing library errors, it can introduce ABI incompatibilities between the LibTorch version used at build time and the one loaded at runtime, leading to symbol resolution errors.

[vboussot]

At the moment, the configuration you tested uses a heavy test preset of the IMPACT metric, which was intentionally designed for stress-testing and validation, not for efficiency on simple mono-modal cases such as MR→MR registration.

This explains the very long runtime you observed. In practice, IMPACT is not intended to replace Elastix’s standard B-spline registration on simple MR→MR cases, where the classical pipeline is already fast and usually sufficient.

I will add lighter, task-specific presets, in particular:

optimized presets for CBCT↔CT registration, and
dedicated presets for MR↔CT registration,

with appropriate model choices, resolutions, and optimization settings. These presets will be much more representative of the intended use cases and will avoid unnecessary computational overhead.