🔄 Slicer IMPACT-Reg

Slicer IMPACT-Reg is an open-source 3D Slicer extension dedicated to multimodal medical image registration.
It integrates the IMPACT similarity metric [1] within the Elastix registration engine, bringing state-of-the-art deep semantic alignment directly into Slicer.

Powered by KonfAI [2], the module provides the following features:

• Fully automated registration pipelines
• GPU-accelerated feature extraction
• Built-in quality assessment and visualization
• Ensemble-based uncertainty quantification

All within a clinically-friendly environment.

---

🖼️ User Interface

IMPACT-Reg registration workflow	Registration evaluation panel
Figure 1 – Multimodal registration interface.	Figure 2 – Evaluation with reference labels.

---

🎥 Demonstration Video

Slicer_IMPACT_Reg_Video.mp4

---

✅ What you can do in 3 minutes (step-by-step tutorial)

This quick tutorial demonstrates the typical clinical workflow: load → run registration → review results → assess reliability.

1) Install and open the module

1. Install 3D Slicer ≥ 5.10
2. Open 3D Slicer and go to Extension Manager
3. Search for ImpactReg and click Install — the KonfAI extension is installed automatically as a required dependency
4. Restart Slicer and open the ImpactReg module from the Registration category
5. The first time you open the module, the impact-reg-konfai Python package (with its matching KonfAI runtime) is installed automatically; the registration presets are then downloaded on demand from Hugging Face

2) Load a case

1. In Slicer, click DICOM (or drag-and-drop a NIfTI / NRRD / MHA file)
2. Load the two volumes: the fixed and the moving image (e.g., Fixed.nii.gz, Moving.nii.gz)
3. Confirm the volumes appears in the Data module and is visible in the slice views

3) Run inference

1. On ImpactReg module go to the Registration tab
2. Select:
   • Fixed volume and Moving volume
   • Preset: the main preset (e.g., IMPACT CT/MR) — it seeds the ensemble and drives evaluation.
   • Ensemble with (optional): add one or more further presets; all selected presets run and their displacement fields are averaged into a single transform.
   • Enable Uncertainty if you want to keep each preset's displacement field for the QA uncertainty step (§5).
3. Click Run
4. Wait for completion: once the process finishes, the warped moving volume is automatically overlaid with the fixed volume in the slice views

✅ You can now inspect the results in 2D and 3D and adjust visualization (opacity).

4) QA with reference (optional)

If a reference annotation (ground truth) is available, it can be:

• an image (e.g., a registered reference volume),
• a segmentation on the fixed and/or moving images,
• or paired landmarks defined in both volumes.

1. Load the reference data (image, segmentation, or landmarks).
2. Go to the Evaluation tab.
3. Select:
   • Images tab for image-based evaluation,
   • Segmentations tab for segmentation-based evaluation,
   • Fiducials tab for landmark-based evaluation.
4. Provide:
   • Fixed and moving ground-truth data,
   • Optional ROI mask,
   • The resulting transform file.
5. Click Run.
6. Review quantitative metrics and qualitative overlays directly inside Slicer.

Generated outputs include:

• MAE_map: voxel-wise Mean Absolute Error (MAE) map between the fixed and warped moving volumes.
• Seg_MAE_map: segmentation-based error map, measuring region-wise discrepancies between corresponding structures.

Reported metrics:

• MAE
• Dice
• TRE

5) QA without reference (uncertainty estimation)

When no ground-truth annotation is available, you can still assess registration reliability from the spread between presets.

1. In the Registration tab, select at least two presets (a main preset + one or more via Ensemble with), enable the Uncertainty checkbox, then Run. Each preset's displacement field is kept in a RegistrationDVFSequence.
2. Go to the Evaluation tab and select No reference (Uncertainty) — the displacement-field sequence produced above is used automatically.
3. Click Run.
4. Review the generated uncertainty maps.

Uncertainty is estimated from the statistical spread of the per-preset displacement fields (multi-preset ensembling).

---

✨ Features

🧠 Deep semantic registration

• IMPACT: feature-space similarity from pretrained segmentation networks
• Multi-preset execution enabling sequential refinement
• GPU or CPU execution
• Optional mask-constrained registration

📊 Built-in evaluation and QA

• Landmark, segmentation, and intensity-based metrics
• Automatic warped volume generation
• 2D/3D synchronized visualization inside Slicer

🔁 Ensemble-based robustness

• Multiple registration presets executed sequentially
• Composite deformation field estimation
• Average transform computation

📉 Uncertainty quantification

• Analysis of the statistical variability of transforms
• Automatic visualization of uncertainty volumes
• JSON metrics export for downstream analysis

---

🧩 Presets & Models

Parameter maps and pretrained models are automatically downloaded from:
📦 VBoussot/ImpactReg on Hugging Face Hub

Each preset includes:

• Parameter maps for Elastix
• Feature extractor models for IMPACT
• A volume-dependent preprocessing function

---

📚 References

1. Boussot, V. et al.
   IMPACT: A Generic Semantic Loss for Multimodal Medical Image Registration.
   arXiv:2503.24121 — 2025

2. Boussot, V. & Dillenseger, J-L.
   KonfAI: A Modular and Fully Configurable Framework for Deep Learning in Medical Imaging.
   arXiv:2508.09823 — 2025