TotalSegmentator CT

TotalSegmentator is a CT dataset built at University Hospital Basel to train a single model that segments most clinically relevant anatomy in any CT scan. Its labels underpin the widely used open-source TotalSegmentator tool, and the data are a common pre-training and benchmark set for multi-organ CT segmentation.

Composition

Version 2.0.1 on Zenodo contains 1,228 CT series, each with one mask file per structure for 117 classes: organs, vessels, bones (including individual vertebrae and ribs), muscles and, new in v2, structures such as the thyroid, prostate, sternum, costal cartilages, kidney cysts and appendicular bones. A metadata table gives age, sex, an anonymized institution letter, study type, scanner, tube voltage, a coarse pathology category and the official train/validation/test split (1,082 / 57 / 89). The series cover every body region, from head CT angiography to whole-body trauma scans. A 102-case subset is offered separately for quick exploration.

Acquisition

Series were drawn at random from the Basel hospital archive for the years 2012, 2016 and 2020, so they reflect routine practice: native and contrast phases, soft-tissue and bone kernels, dual-energy scans and a wide mix of pathologies. Most images come from Siemens scanners, with smaller numbers from Philips and GE. All images were resampled to 1.5 mm isotropic resolution.

Annotations

Labels were produced iteratively. Existing models and atlas tools gave first drafts, two physicians reviewed and corrected them, and an nnU-Net retrained on the corrected cases produced better drafts for the next round, until every case had been manually reviewed. Version 2 also fixed systematic errors in classes such as the femur, hip, heart, aorta, liver, spleen and kidneys.

Known limitations

• The v1 paper treats each series as one patient; the v2 release does not state this, so counts here are given as series.
• The paper describes v1 (1,204 series, 104 classes); v2 changed the test split size and the class list.
• Most data come from one hospital and one vendor.
• Pathology information is missing for many series and is only a coarse category.