Third-party material redistributed in this repository
=====================================================

Every model here is a TorchScript export of a pretrained network published by someone else. The
weights are unchanged; what this repository adds is a wrapper that normalises the input from the
statistics IMPACT passes and returns one tensor per feature layer. The export scripts are published in
https://github.com/vboussot/ImpactLoss under Data/Models/builds/.

Anatomix/Anatomix.pt, Anatomix/AnatomixDevViT.pt
------------------------------------------------
  Weights:  https://huggingface.co/neeldey/anatomix (anatomix.pth, anatomix-dev-vit.pth)
  Code:     https://github.com/neel-dey/anatomix
  License:  MIT, Copyright 2024 Neel Dey -- LICENSES/anatomix-MIT.txt
  Paper:    Dey et al., Learning General-purpose Biomedical Volume Representations using Randomized
            Synthesis, ICLR 2025, arXiv:2411.02372

AnatomixDevViT additionally builds on PrimusV2 from dynamic-network-architectures, which anatomix
configures and extends (output normalisation, register-token initialisation, QK normalisation):

  Code:     https://github.com/MIC-DKFZ/dynamic-network-architectures
  License:  Apache License 2.0, Copyright 2022 Division of Medical Image Computing, German Cancer
            Research Center (DKFZ) -- LICENSES/dynamic-network-architectures-Apache-2.0.txt

TS/M258.pt, TS/M291-M295.pt, TS/M297-M298.pt, TS/M730-M733.pt, TS/M850-M853.pt
-------------------------------------------------------------------------------
  Weights:  TotalSegmentator, https://github.com/wasserth/TotalSegmentator
  License:  Apache License 2.0 -- LICENSES/Apache-2.0.txt
  Paper:    Wasserthal et al., TotalSegmentator: robust segmentation of 104 anatomic structures in CT
            images, Radiology: Artificial Intelligence, 2023, arXiv:2208.05868

  TotalSegmentator publishes its tasks under two different terms. The models here come from the tasks
  it offers under Apache 2.0: `total` (M291-M295, M297, M298), `total_mr` (M730-M733 in the 2.2/2.4
  series, M850-M853 since) and `lung_vessels` (M258). Its other tasks -- heartchambers_highres,
  tissue_types, brain_structures, appendicular_bones, face, thigh_shoulder_muscles, coronary_arteries
  and the rest of that list -- need a license number from the authors, free for non-commercial use and
  paid otherwise, and their weights must NOT be redistributed here.

MRSeg/MRSeg.pt
--------------
  Weights:  MRSegmentator, https://github.com/hhaentze/MRSegmentator
  License:  Apache License 2.0 -- LICENSES/Apache-2.0.txt
  Paper:    Haentze et al., MRSegmentator: Multi-Modality Segmentation of 40 Classes in MRI and CT,
            2024, arXiv:2405.06463

SAM2.1/SAM2.1_Small.pt, SAM2.1/SAM2.1_Tiny.pt
----------------------------------------------
  Weights:  SAM 2, https://github.com/facebookresearch/sam2
  License:  Apache License 2.0, Copyright (c) Meta Platforms, Inc. and affiliates --
            LICENSES/Apache-2.0.txt
  Paper:    Ravi et al., SAM 2: Segment Anything in Images and Videos, 2024, arXiv:2408.00714

Dino/DinoV2_Small.pt
--------------------
  Weights:  DINOv2, https://github.com/facebookresearch/dinov2
  License:  Apache License 2.0, Copyright (c) Meta Platforms, Inc. and affiliates --
            LICENSES/Apache-2.0.txt
  Paper:    Oquab et al., DINOv2: Learning Robust Visual Features without Supervision, 2023,
            arXiv:2304.07193

  DINOv3 is NOT redistributed here: its checkpoints fall under the DINOv3 License Agreement rather than
  an open-source license. The export script in ImpactLoss builds them locally from your own download.

VGG/VGG16.pt
------------
  Weights:  torchvision, https://github.com/pytorch/vision
  License:  BSD 3-Clause, Copyright (c) Soumith Chintala 2016 --
            LICENSES/torchvision-BSD-3-Clause.txt
  Paper:    Simonyan and Zisserman, Very Deep Convolutional Networks for Large-Scale Image
            Recognition, 2014, arXiv:1409.1556

MIND/*.pt
---------
  No third-party weights: MIND is a handcrafted descriptor, implemented here from its paper.
  Paper:    Heinrich et al., MIND: Modality independent neighbourhood descriptor for multi-modal
            deformable registration, Medical Image Analysis, 2012, doi:10.1016/j.media.2012.05.008