> ## Content Index
> Fetch the complete content index at: https://globalfeed.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# TotalSegmentator now predicts weight, height, age and sex from CT and MR scans: 20 seconds on CPU
- URL: https://globalfeed.ai/en/totalsegmentator-now-predicts-weight-height-age-and-sex-from-ct-and-mr-scans-20-seconds-on-cpu/
- Published: 2026-09-04T23:59:53.000Z
- Updated: 2026-09-05T00:02:27.000Z
- Description: The open-source tool from a University Hospital Basel team fills in patient details directly from the image, for archives where that metadata is missing or unreliable. On CT the mean weight error is 3.90 kg, height 3.68 cm, sex F1 0.990. No GPU required. The developers state plainly that this is not
- Author: GlobalFeed Editor
- Tags: TotalSegmentator, medical imaging, open source, AI, CT, MRI, x-JakobWasserthal, dil-en, elle, video, x-gitti

Hospital imaging archives have a quiet problem: the patient and acquisition details stored alongside a CT or MR scan are often missing, and sometimes wrong. Those details matter for clinical decisions, image quality control and automated research pipelines.

A team centred on University Hospital Basel proposes filling the gap from the image itself. A new task added to the widely used open-source segmentation tool **TotalSegmentator** predicts patient and acquisition characteristics by looking at the scan. The work was [published on arXiv on 29 August 2026](https://arxiv.org/abs/2608.29348?ref=globalfeed.ai).

## What it predicts

For both modalities: **weight, height, age, sex**, contrast presence, vertebral coverage and image noise.

For CT only: scanner manufacturer, tube voltage, tube current, convolution kernel and post-injection time. For MR only: sequence class.

## The numbers

The model consists of separately trained **3D ResNet-10 ensembles**, trained on **57,291 CT and 43,200 MR clinical examinations** spanning 2011 to 2025.

On the internal test set, CT results: mean absolute error of **3.90 kg** for weight, **3.68 cm** for height, **4.42 years** for age, and sex F1 of **0.990**. For MR, in the same order: 4.34 kg, 4.62 cm, 7.13 years and 0.970.

The external dataset result is the more meaningful one, because it shows whether the model holds up outside its own hospital. On the external CT set: weight **4.45 kg**, height **4.05 cm**, age **5.17 years**, sex F1 **0.971**. There is a drop, but a small one; the model does not appear to have overfitted to its own institution.

## Why it needs no GPU

This is the detail that matters most in practice. Inference on CPU takes **20 seconds for CT and 12 seconds for MR**. You do not need to buy a GPU server to run it, which makes sweeping an existing archive of old examinations a job an ordinary server can do.

## How it runs

The tool is called from the command line in a single line:

```
totalseg_get_body_stats -i ct.nii.gz -o body_stats.json -m ct
```

The task additionally requires the `timm` and `monai` packages. For anyone who would rather not install anything to try it, the team has also put up a web interface: [compute.totalsegmentator.com/body-stats](https://compute.totalsegmentator.com/body-stats/?ref=globalfeed.ai).

## Licensing: the core is open, some models are not

TotalSegmentator itself is licensed **Apache-2.0**, which permits commercial use. But per the repository's own notes, several specialised models are separately licensed: non-academic use means contacting the developer. Before deploying inside an organization, check which bucket the specific task you need falls into.

## And the actual warning

The developers write it plainly in the repository: *"This is not a medical device and is not intended for clinical usage."*

That sentence is not cosmetic. The tool's stated purpose is completing missing metadata, quality control and feeding research pipelines. Reading a patient's weight or age from here for a clinical decision is not the job it was designed for. In an enterprise deployment that distinction needs to be written down, because once the output arrives as a JSON file, the limits of the model that produced it become invisible.

## Keeping it in proportion

One reminder points the other way. A model that can infer age, sex and body measurements from an image raises a separate question for datasets believed to be de-identified: even with patient details stripped, the image itself still carries part of that information. The paper does not open this discussion, nor is that its purpose, but an institution about to open its archive to research should account for it.

## Sources

- **Primary source, the paper:** Jakob Wasserthal, Joshy Cyriac, Michael Bach, Kimia Mozahheb Yousefi, Minh-Son To, Mate Sik, Cedric Hemon, Thomas Weikert, Martin Segeroth, *"Extending TotalSegmentator: Predicting Patient and Acquisition Characteristics from CT and MR Images"*, arXiv:2608.29348, 29 August 2026: [arxiv.org/abs/2608.29348](https://arxiv.org/abs/2608.29348?ref=globalfeed.ai)
- **Code, licence and usage:** the TotalSegmentator repository, from which the command and licence notes are taken: [github.com/wasserth/TotalSegmentator](https://github.com/wasserth/TotalSegmentator?ref=globalfeed.ai)
- **Web interface:** [compute.totalsegmentator.com/body-stats](https://compute.totalsegmentator.com/body-stats/?ref=globalfeed.ai)
- **Announcement:** posted by Jakob Wasserthal, the paper's first author and the tool's developer, on his own account, 4 September 2026.