TS-Portable models

ONNX models for TS-Portable, a self-contained program that runs TotalSegmentator on the CPU. They are the seven models of the total task of TotalSegmentator version 2.12.0 (the five 1.5-mm submodels, the 3-mm model, and the 6-mm localization model), converted to ONNX, and an 8-bit integer (INT8) version of each.

These files are the weights only. To segment a CT scan, use the TS-Portable program from the GitHub repository; its README walks through the Windows download and the command line.

Files

File	Contents	Size
ts-portable-models.zip	INT8 models (*_int8.onnx), model cards, label map, LICENSE, NOTICE, SHA256SUMS.txt, in a models/ folder	172 MB (212 MB unpacked)
fp32/	the same seven models unquantized (*.onnx), model cards, label map, LICENSE, NOTICE, SHA256SUMS.txt	757 MB
LICENSE, NOTICE	Apache License 2.0 and the attribution and change notice for the weights

• INT8 (ts-portable-models.zip) is what TS-Portable uses by default (--device openvino --int8). It needs the OpenVINO execution provider of ONNX Runtime (onnxruntime-openvino 1.24.1) on an x86-64 CPU; it does not run on macOS. Unzip it next to the program, so that models/ sits beside the executable.
• fp32 (fp32/) runs through any ONNX Runtime provider: CPU (including macOS), CUDA on NVIDIA GPUs (--device cuda), or DirectML on Windows (--device directml). Point TS-Portable at it with --models-dir fp32.

Each model card (*.model_card.json) records the nnU-Net configuration the program needs: target spacing, patch size, axis order, intensity normalization, resampling settings, and labels. total_label_map.json maps structure names to TotalSegmentator's label numbers.

How the INT8 models were made

Accuracy-controlled post-training quantization with NNCF 3.2.0 (quantize_with_accuracy_control): per-channel INT8 weights and UINT8 activations, with any operation whose quantization would cost more than the accuracy budget restored to fp32 (maximum patch-fidelity drop 0.005, at most four restoration iterations). The first and last convolution of each model stay in fp32. Calibration, selection and held-out patch testing used CT-ORG volumes 0–27, which are disjoint from the evaluation data. The script and the exact settings are in the GitHub repository (scripts/quantize_nncf_sequential.py, docs/METHOD.md).

Evaluation

In the accompanying study, TS-Portable with these INT8 models was compared with the distributed PyTorch TotalSegmentator package on 100 external CT scans of the AMOS validation set, 15 abdominal organs, at 1.5 mm:

• Accuracy: noninferior within a 0.01 margin for per-case mean Dice (difference 0.00017; 95% CI, 0.00003–0.00031) and 3-mm normalized surface distance (difference 0.00015; 95% CI, −0.00002 to 0.00031).
• Speed: on a 12-core workstation CPU (AMD Ryzen Threadripper PRO 7945WX), the median time per scan was 48.0 s, against 342.5 s for the reference implementation on the same CPU and 61.5 s on an NVIDIA RTX A6000 GPU.
• fp32 models: the unquantized ONNX models run by the same program differed from the reference implementation by at most 0.00012 in mean Dice or surface distance.

The per-case results and the code behind every number are in the repository's evaluation/ folder.

Intended use and limitations

• For research. TS-Portable is not a medical device and has no regulatory clearance.
• The models are TotalSegmentator's: they reproduce its correct segmentations and its errors. Check masks on the source images when a result looks implausible.
• INT8 results depend on the CPU and runtime. The paper's results used ONNX_TOTALSEG_OPENVINO_INFERENCE_PRECISION=f32; without it, CPUs with BF16 support may give slightly different output.

Checksums

b8a20812f00e13192cb5bc3912b902950a74e91756a275311230691030bfbc8e  ts-portable-models.zip

Per-file checksums are in models/SHA256SUMS.txt inside the zip and in fp32/SHA256SUMS.txt; the INT8 ones are also in the GitHub repository (deployment/models.sha256).

License and attribution

Apache License 2.0 (LICENSE; a copy is also inside the zip and in fp32/). The weights are converted and quantized from the total task of TotalSegmentator (Apache-2.0), which is built on nnU-Net (Apache-2.0); NOTICE gives the attribution and the changes made. Please cite TotalSegmentator:

Wasserthal J, Breit HC, Meyer MT, et al. TotalSegmentator: robust segmentation of 104 anatomic structures in CT images. Radiol Artif Intell 2023; 5:e230024.

and the TS-Portable paper:

Lee JH. Quantized TotalSegmentator for Fast Abdominal Organ Segmentation on CPUs. medRxiv 2026. doi:10.64898/2026.10.23.26366189