Valentin Boussot

1. 1

   Multimodal registration · IMPACT-Reg

   A modality-agnostic semantic similarity metric built from pretrained segmentation features, where intensity-based measures break down across modalities. The registration engine is unchanged (only the similarity metric changes), so it plugs into both algorithmic (Elastix) and learning-based registration.

   • IMPACT-Reg
   • Elastix

   What problem does this solve?

   Problem

   Aligning two scans of the same anatomy across modalities is hard: their intensities don't correspond, so the intensity-based measures classic registration relies on break down.

   Key idea

   Compare images in the feature space of pretrained segmentation networks, not raw voxels. A simple L1/L2 distance there becomes a modality-agnostic similarity measure.

   Two modes

   In Elastix it runs in a Jacobian mode (features recomputed each iteration for the most faithful gradient) or a lighter Static mode using precomputed features. The best backbone and layer depend on the anatomy and modality: no universal recipe.

   Result

   1.20 mm median TRE on CT/CBCT, ranked 1st on Learn2Reg Lung CT and 2nd on Abdomen MR–CT, and merged into the official Elastix toolkit. It ships in the IMPACT ecosystem (below) as IMPACT-Reg.

   

2. 2

   Anatomy-preserving CT synthesis · IMPACT-Synth

   Cross-modality CT-like synthesis as domain adaptation under an anatomy-preservation constraint: a content / style split that makes CBCT and MRI usable by CT-trained tools without displacing, erasing or inventing patient anatomy.

   • IMPACT-Synth
   • Content / style
   • SAM perceptual loss
   • TotalSegmentator

   What problem does this solve?

   Problem

   Many imaging tools are trained only on CT, so CBCT and MRI can't use them. Making those scans CT-like fixes that, but naive translation can hallucinate or displace anatomy.

   Key idea

   Separate content from style: hold the anatomy to the source image, move only the appearance toward CT. Anatomy is compared in the same SAM / TotalSegmentator feature space, so the constraint is anatomical, not voxel-wise.

   Why it matters

   Relaxing the rigid voxel-to-voxel target is what cuts the geometric bias that, as the next section shows, supervised synthesis otherwise inherits. A SAM perceptual loss (matching structures in a Segment Anything feature space, not intensities) is complementary: it reduces the blur of voxel-wise regression and improves downstream segmentation, but, still trained on the registered pairs, it does not remove the systematic bias on its own.

   Result

   On VATSop, CBCT→CT-like MAE fell from 227.6 to 129.7 HU and lobe-segmentation Dice rose from 0.898 to 0.930. Its public form placed 3rd at SynthRAD 2025 (Tasks 1 & 2), and ships in the IMPACT ecosystem (below) as IMPACT-Synth.

   

3. 3

   Statistical deformation model · wPCA

   A weighted-PCA prior over lung deformation, learned across the cohort and recovered for a new patient as a latent inverse problem from the pleural surface alone, predicting the change of pose so one intra-operative CBCT can be dropped.

   • wPCA
   • Statistical shape model
   • Latent inverse problem

   What problem does this solve?

   Problem

   The reference pipeline needs a second, intra-operative CBCT just to recover how the lung shifts when the patient is repositioned, before it is deflated: extra dose, time and complexity.

   

   Key idea

   Learn a compact deformation prior across the cohort. Every patient is registered into a common template space, where a weighted PCA captures the plausible pose changes as a mean field plus a few dominant modes.

   Recovery

   For a new patient, recover those modes' coefficients as a latent inverse problem from the pleural surface alone, then reconstruct the dense, patient-specific field by B-spline interpolation.

   Result

   Median localisation error drops from 11.96 to 6.33 mm (about 47% over a rigid baseline), removing one intra-operative CBCT. It has no public release, so there is no repository link.

   

• 1st placeLearn2Reg · Lung CT
• 2nd placeLearn2Reg · Abdomen MR–CT
• 6th placeLearn2Reg · Thorax CBCT
• 3rd placeSynthRAD 2025 · Tasks 1 & 2
• 2nd placePANTHER · Task 2
• 3rd placeCURVAS
• 3rd placeCURVAS-PDACVI
• 6th placeTrackRAD
• 10th placeSynthRAD 2023 · Task 2

• 

  Registration metric

  IMPACT-Reg

  A modality-agnostic similarity metric, shipped as the reusable ImpactLoss library, that drops into a standard registration pipeline.

  ImpactLoss repository (opens in a new tab)
• 

  CT synthesis

  IMPACT-Synth

  Open training code and models for cross-modality CT synthesis, released in the public, reproducible form that placed 3rd at SynthRAD 2025.

  Synthrad2025_Task_1 repository (opens in a new tab)

Also ships ImpactElastix repository (opens in a new tab) · impact-torchscript-models on Hugging Face (opens in a new tab)

This is a selection. For the complete, up-to-date list see Google Scholar (opens in a new tab).