File size: 4,298 Bytes
5e39b69 0e736fa 5e39b69 0e736fa 5e39b69 0e736fa 5e39b69 0e736fa | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 | 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
|