🧠 TorchScript Models for the IMPACT Semantic Similarity Metric

This repository provides a collection of TorchScript-exported pretrained models designed for use with the IMPACT similarity metric, enabling semantic medical image registration through feature-level comparison.

The IMPACT metric is introduced in the following preprint, currently under review:

IMPACT: A Generic Semantic Loss for Multimodal Medical Image Registration
V. Boussot, C. Hémon, J.-C. Nunes, J. Dowling, S. Rouzé, C. Lafond, A. Barateau, J.-L. Dillenseger
arXiv:2503.24121 [cs.CV]

🔧 The full implementation of IMPACT, along with its integration into the Elastix framework, is available in the repository:
➡️ github.com/vboussot/ImpactLoss

This repository also includes example parameter maps, TorchScript model handling utilities, and a ready-to-use Docker environment for quick experimentation and reproducibility.

Anatomix variants

Two exports of the anatomix encoders:

File	Network	Feature channels	Input size
`Anatomix/Anatomix.pt`	U-Net, 4 levels (`anatomix.pth`)	16	any (multiple of 16 per axis)
`Anatomix/AnatomixDevViT.pt`	PrimusV2 ViT, 27.1 M parameters (`anatomix-dev-vit.pth`)	32	128 x 128 x 128 only

The ViT's positional embedding is sized for a 128-voxel cube, so it has to be scored patch by patch at that size: PatchSize 128 128 128 for Elastix, feature_patch: 128 for the KonfAI FireANTs engine.

Both are traced with their network in evaluation mode, so their normalisation layers use the learnt statistics, and wrapped in a scripted module, so nb_layers, stats and direction keep their default values and a caller may pass only the image. Anatomix.pt was previously traced with the network in training mode, where its BatchNorms read each batch's own statistics: its features spanned roughly +-109 against +-5 now, and the two correlate at 0.21 on random input. Pin the revision before 2026-09-23 to reproduce a result obtained with it:

ImpactModelConfiguration("VBoussot/impact-torchscript-models:Anatomix/Anatomix.pt", revision="<sha>")

Dino/DinoV2_Small.pt is exported the same way, for the same reason: traced with an empty stats argument, its normalisation branch was frozen on the image's own min and max, so the statistics IMPACT passes were ignored. Scored with statistics, the two exports differ by about 7e-3; without them they agree to 9e-6.

NOTICE names, for every model here, where its weights come from and under which terms, and LICENSES/ carries those licenses in full. Two of them bound what may be added to this repository: TotalSegmentator publishes part of its tasks under Apache 2.0 and the rest behind a license number, of which only the open ones are exported here; and DINOv3's checkpoints are covered by the DINOv3 License Agreement, so they are built locally rather than redistributed.

The export scripts live in ImpactLoss/Data/Models/builds/Anatomix.

📚 Pretrained Model

The TorchScript models provided in this repository were exported from publicly available pretrained networks. These include:

TotalSegmentator (TS) — U-Net models trained for full-body anatomical segmentation
MRSegmentator (MRSeg) — U-Net models trained for full-body anatomical segmentation in MRI and CT
Segment Anything 2.1 (SAM2.1) — Foundation model for segmentation on natural images
DINOv2 — Self-supervised vision transformer trained on diverse datasets
Anatomix — Transformer-based model with anatomical priors for medical images

Each model provides multiple feature extraction layers. This can be configured through the LayerMask parameter in the IMPACT configuration.

In addition, the repository also includes:

MIND — A handcrafted descriptor, wrapped in TorchScript

Model	Specialization	Paper / Reference	Field of View	License	Preprocessing
MIND	Handcrafted descriptor	Heinrich et al., 2012	`2rd + 1` (r: radius, d: dilation)	Apache 2.0	Normalize intensities to [0, 1]
SAM2.1	General segmentation (natural images)	Ravi et al., 2023	29	Apache 2.0	Normalize intensities to [0, 1], then standardize with mean 0.485 and std 0.229
TS Models	CT/MRI segmentation	Wasserthal et al., 2022	`2^l + 3` (l: layer number)	Apache 2.0 (open tasks only, see `NOTICE`)	Canonical orientation for all models. For MRI models (e.g., TS/M730–M733-M850–M853), standardize intensities to zero mean and unit variance. For CT models (e.g., TS/M258, TS/M291), clip intensities + normalize model dependant
MRSegmentator	CT/MRI segmentation	Häntze et al., 2024	`2^l + 3` (l: layer number)	Apache 2.0	Standardize intensities to zero mean and unit variance.
Anatomix	Anatomy-aware encoder (U-Net, 16 features)	Dey et al., 2024	Global(Static mode)	MIT	Normalize intensities to [0, 1]
AnatomixDevViT	Anatomy-aware encoder (ViT, 32 features)	Dey et al., 2024	128 (fixed input)	MIT + Apache 2.0	Normalize intensities to [0, 1]
DINOv2	Self-supervised vision transformer	Oquab et al., 2023	14	Apache 2.0	Normalize intensities to [0, 1], then standardize with mean 0.485 and std 0.229