New AI Framework Segments 167 Body Structures Across 22,000 CT Scans

Researchers have introduced CADS, a unified data-and-model framework for whole-body anatomical segmentation in computed tomography, published as an open-access article in Nature Communications on 28 October 2026. The work addresses a persistent problem in medical imaging AI: most existing segmentation methods remain organ-specific, built as fragmented models with inconsistent label definitions and heterogeneous training and evaluation protocols that make large-scale quantitative image analysis difficult.

At the core of the framework is a new anatomical dataset. The team constructed a SNOMED-CT-referenced label space that harmonizes heterogeneous imaging cohorts drawn from more than 100 imaging centers. The resulting resource provides complete annotations of 167 anatomical structures across 22,022 CT volumes, which the authors describe as markedly expanding both the anatomical coverage and the scale available compared with existing segmentation resources.

Because real-world CT scans are often only sparsely annotated, the researchers developed a strategy to complete missing labels. They combined self-training with a shape-informed quality control step based on neural implicit functions, allowing the pipeline to curate high-quality pseudo-labels on scans that lack full manual annotation. This curation step is what turns the heterogeneous, partially labeled source cohorts into a consistently labeled training resource.

Trained on this curated dataset, the CADS segmentation model showed robust whole-body performance across distribution shifts and improved performance relative to state-of-the-art approaches when evaluated across 18 public datasets. The team further validated the model in an independent, expert-reviewed real-world oncology cohort of 2,864 CT scans, providing an external check on performance beyond the public benchmark datasets.

Alongside the dataset and trained models, the researchers released user-friendly software, which they position together as an accessible and reproducible foundation for large-scale anatomical segmentation and computational anatomy research. The article is currently shared as an early version of the peer-reviewed, accepted research, carrying a permanent DOI and subject to further edits before the final Version of Record.

Subject of Research: A unified dataset and AI model for whole-body anatomical segmentation in computed tomography

Article Title: CADS: a comprehensive anatomical dataset and segmentation for whole-body anatomy in computed tomography

Article References: Xu, M., Navarro, F., Fritsak, M., Hamamci, I. E., Shit, S., Wittmann, B., Er, S., Christ, S. M., de la Rosa, E., Deseoe, J., Graf, R., Möller, H., Sekuboyina, A., Peeken, J. C., Becker, S. S., Baldini, G., Haubold, J., Nensa, F., Hosch, R., … Menze, B. (2026). CADS: a comprehensive anatomical dataset and segmentation for whole-body anatomy in computed tomography. Nature Communications. https://doi.org/10.1038/s41467-026-78732-1

Image Credits: AI Generated

DOI: 10.1038/s41467-026-78732-1

Keywords: computed tomography, image segmentation, anatomical dataset, deep learning, SNOMED-CT, pseudo-labels, oncology imaging, medical imaging AI, whole-body anatomy, computational anatomy, CADS, comprehensive