segment-anatomy

name	segment-anatomy
description	Workflow for segmenting anatomical structures and measuring their volumes on CT or MR with TotalSegmentator

Anatomy segmentation & volumetry workflow

segment_anatomy labels anatomical structures in a CT or MR volume and can write a per-structure volume CSV. The default task total covers 117 structures on whole-body CT; total_mr is the 50-structure MR counterpart; around 30 focused tasks cover anatomy the whole-body models miss or segment less well.

When to use

• The user wants organs, bones, muscles or vessels labelled on a CT or MR volume.
• The user wants a volume or a measurement for a specific structure (liver volume, spleen volume, muscle area).
• The user wants a mask to feed another tool (cropping, radiomics, dose planning).

Steps

1. Resolve anatomy to a task before running anything. If the user named a structure rather than a task, call find_structures first. Several tasks segment the same structure at different scope: the liver appears in total, but Couinaud segments need liver_segments and lesions need liver_lesions. Picking the general model when the user wanted the specific one produces a confident, useless answer. list_segmentation_tasks shows everything available.
2. Match the modality. CT tasks and MR tasks are not interchangeable — a task name ending in _mr is the MR variant. segment_anatomy checks the volume's intensity range and returns a warnings entry if it looks like the wrong modality, but it does not block: relay the warning and confirm before accepting the result.
3. Run segment_anatomy with device="auto". Check the resolved device in the result — if it fell back to cpu, say so, because runtime changes from about a minute to many minutes.
4. Report the requested structure. When the user asked for a specific volume, read the CSV at volumes_path and report the matching rows (values in mm³). Offer the full CSV rather than dumping every row.

Narrowing a whole-body run

On total / total_mr only, pass structures=[...] to segment just what was asked for. It crops the volume to that region first, so it is substantially faster than a full run — the right default when the user wants one organ. Names must match the task's spelling exactly (kidney_left, not "left kidney"); list_task_structures gives the exact list. On any other task structures is rejected: those tasks are already specific.

Gotchas

• DICOM input needs converting first. These tools take NIfTI. If the user points at a DICOM series, convert it with the DICOM stack before segmenting.
• speed trades away small structures. "fast" (3 mm) and "fastest" (6 mm) are for triage and for CPU runs. Anything small — adrenal glands, vessels, nodules — needs "standard". Never quote a volume measured at "fastest" as a real number without saying which resolution produced it.
• Output is one multilabel volume by default (*_dseg.nii.gz), which the workspace viewer renders directly as a labelled overlay — offer to open it. The label index → name mapping is the CSV at labels_path. Only pass separate_masks=True when a downstream tool actually needs one file per structure.
• Comparing volumes across subjects — raw mm³ scales with body size. For cohort comparisons normalise (e.g. by height, or by a reference structure) and say which normalisation you used.
• Some structures need a license and are not installed. find_structures reports those under unavailable with the reason (body-composition tissue types, brain structures, coronary arteries, face masking, and the non-commercial aneurysm model). Relay that verbatim — do not substitute a different task and present it as the thing that was asked for.
• These are estimates, not clinical findings. TotalSegmentator is not a medical device. Present output as a research measurement, and never as a diagnosis.