TAMIAS mirror of the Zenodo liver-segmentation models (ONNX)
ONNX exports (opset 17, fp32, fixed patch) and TAMIAS manifests converted from:
- LightningMedSeg3D trained weights: 10.5281/zenodo.21037952. The weights are CC-BY-4.0, see
LICENSE-zenodo-21037952. The architecture code, LightningMedSeg3D, is AGPL-3.0-or-later. - nnU-Net v2 Dataset006_Liver (liver + lesion, LiTS 2017, from BAMF Health): 10.5281/zenodo.11582728. CC-BY-4.0. nnU-Net is Apache-2.0.
Conversion code: scripts/zenodo/ in ArioMoniri/semikap. The GitHub release is zenodo-models-v1.
zenodo-models-index.json (schema tamias.model-index.v1) lists every file with its sha256, labels, parity and licence.
Input contract
- Input tensor: float32
[1,1,Z,Y,X]after reorienting tomanifest.orientation(RAS) and resampling tomanifest.spacing. - Output tensor: logits
[1,C,Z,Y,X]. - LightningMedSeg3D models: the manifest applies a
windownormalisation (level 37.5, width 425 HU). - nnU-Net model: CTNormalization is baked into the graph, so the manifest uses
none.
Parity check: PyTorch runs the reference pipeline and onnxruntime runs the TAMIAS pipeline. Both run on a random-HU patch and a synthetic CT phantom.
| id | trained on | classes | patch (x,y,z) | spacing mm | normalization | ONNX MB | parity max abs diff / argmax agreement | status |
|---|---|---|---|---|---|---|---|---|
lms3d_attention_unet | BTCV | 14 | [96, 96, 96] | [1.5, 1.5, 2.0] | window | 162.7 | 2.15e-02 / 1.0000 | ok |
lms3d_medformer | BTCV | 14 | [96, 96, 96] | [1.5, 1.5, 2.0] | window | 155.2 | 3.60e-02 / 1.0000 | ok |
lms3d_resunet | BTCV | 14 | [96, 96, 96] | [1.5, 1.5, 2.0] | window | 162.3 | 6.82e-03 / 1.0000 | ok |
lms3d_segformer | BTCV | 14 | [96, 96, 96] | [1.5, 1.5, 2.0] | window | 18.1 | 3.24e-05 / 1.0000 | ok |
lms3d_swin_unetr | BTCV | 14 | [96, 96, 96] | [1.5, 1.5, 2.0] | window | 143.4 | 5.06e-03 / 1.0000 | ok |
lms3d_unet | BTCV | 14 | [96, 96, 96] | [1.5, 1.5, 2.0] | window | 65.1 | 6.02e-03 / 1.0000 | ok |
lms3d_unetpp | BTCV | 14 | [96, 96, 96] | [1.5, 1.5, 2.0] | window | 17.9 | 5.66e-02 / 1.0000 | ok |
lms3d_unetr | BTCV | 14 | [96, 96, 96] | [1.5, 1.5, 2.0] | window | 371.3 | 5.90e-03 / 1.0000 | ok |
lms3d_vnet | BTCV | 14 | [96, 96, 96] | [1.5, 1.5, 2.0] | window | 182.6 | 1.82e-02 / 1.0000 | ok |
nnunet_liver_lits | LiTS 2017 | 10 | [128, 128, 128] | [0.767578125, 0.767578125, 1.0] | zscore_volume | 124.8 | 1.47e-02 / 1.0000 | ok |
nnunet_liver_lits_f1 | LiTS 2017 | 10 | [128, 128, 128] | [0.767578125, 0.767578125, 1.0] | zscore_volume | 124.8 | 5.94e-03 / 1.0000 | ok |
nnunet_liver_lits_f2 | LiTS 2017 | 10 | [128, 128, 128] | [0.767578125, 0.767578125, 1.0] | zscore_volume | 124.8 | 5.24e-03 / 1.0000 | ok |
nnunet_liver_lits_f3 | LiTS 2017 | 10 | [128, 128, 128] | [0.767578125, 0.767578125, 1.0] | zscore_volume | 124.8 | 1.22e-02 / 1.0000 | ok |
nnunet_liver_lits_f4 | LiTS 2017 | 10 | [128, 128, 128] | [0.767578125, 0.767578125, 1.0] | zscore_volume | 124.8 | 2.14e-03 / 1.0000 | ok |
Research use only. Not a medical device. Do not use for clinical decisions.
Citation
@dataset{fdezgonzalez_lightningmedseg3d_weights_2026,
author = {Fdez-Gonz{\'a}lez, Marcos and Nodar-Corral, Lois and Fdez-Vidal, Xose R. and
Est{\'e}vez-Fern{\'a}ndez, Sergio and Comesa{\~n}a, Enrique},
title = {Trained Weights of Nine 3D Medical Image Segmentation Networks on {BTCV} and
{MSD} {Task03} ({Liver}) --- {LightningMedSeg3D}},
year = {2026}, publisher = {Zenodo}, doi = {10.5281/zenodo.21037952}
}
@software{lightningmedseg3d,
author = {Fdez-Gonz{\'a}lez, Marcos and Removirt},
title = {{LightningMedSeg3D}: A {PyTorch} {Lightning} framework for 3D medical image segmentation},
url = {https://github.com/Removirt/LightningMedSeg3D}
}
@software{murugesan_liver_nnunet_2024,
author = {Murugesan, Gowtham Krishnan and Van Oss, Jeff and McCrumb, Diana},
title = {Pretrained model for 3D semantic image segmentation of the liver and liver lesions from ct scan},
year = {2024}, publisher = {Zenodo}, doi = {10.5281/zenodo.11582728}
}
@article{isensee2021nnunet,
author = {Isensee, F. and Jaeger, P. F. and Kohl, S. A. and Petersen, J. and Maier-Hein, K. H.},
title = {nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation},
journal = {Nature Methods}, volume = {18}, number = {2}, pages = {203--211}, year = {2021}
}
@article{bilic2023lits,
author = {Bilic, P. and Christ, P. and Li, H. B. and others},
title = {The Liver Tumor Segmentation Benchmark ({LiTS})},
journal = {Medical Image Analysis}, volume = {84}, pages = {102680}, year = {2023}
}