CloudSegmentator

CloudSegmentator runs AI segmentation models on images from the NCI Imaging Data Commons (IDC) on cloud VMs, and converts the results to standard DICOM that can be shared back through IDC:

• Segmentations of anatomy, as DICOM Segmentation (SEG) objects with each structure coded in SNOMED CT.
• Radiomics features of each segmented structure, as JSON and as DICOM Structured Reports (TID 1500).

The models are TotalSegmentator and MOOSE. The workflows are written in WDL, run on Terra, and are published on Dockstore. They read public IDC data directly, or DICOM from a private Google Cloud Storage bucket.

Results

When	Models and workflow	Images	Results
2023–24	TotalSegmentator; older workflow, release v1.2.0	126,088 CT series from NLST	In IDC as TotalSegmentator-CT-Segmentations
Oct 2026	TotalSegmentator and MOOSE; harmonized workflow	3,059 CT series from CMB and CPTAC	On Zenodo, pending ingestion into IDC

Which workflow to use

For new runs, use the harmonized workflow. One WDL runs every supported model, and only the inference step is model-specific, so adding a model takes one notebook and one Docker image.

Dockstore workflows	Models	Status	Docs
SegmentatorTwoVmWorkflowOnTerra	TotalSegmentator 2.18.0 (total, lung_vessels); MOOSE (10 clin_ct_* models)	Current. Validated on Terra in Sep 2026; used for the 2026 run	Harmonized
TotalSegmentator*VmWorkflowOnTerra	TotalSegmentator 1.5.6	Older; used for the NLST run	TotalSegmentator
Moose*VmWorkflowOnTerra	MOOSE	Older; replaced by the harmonized workflow	MOOSE

The older workflows stay in the repository and on Dockstore so that past runs can be reproduced. Open issues with the harmonized workflow, such as its untested import into a DICOM store, are listed under Known gaps.

How the harmonized workflow works

Each workflow instance processes one batch of series on two VMs. Every step is a Jupyter notebook, run by papermill inside a Docker image.

GPU VM (preemptible)                         CPU VM
  convert     DICOM → NIfTI (dcm2niix)         output conversion
  inference   NIfTI → label maps        ──▶      NIfTI → DICOM SEG (dcmqi)
              (one notebook and one              radiomics (pyradiomics or Radiomics.jl)
               image per model)                  DICOM SR (TID 1500)
                                                 optional upload to GCS and a DICOM store

Finished work is checkpointed to a bucket, so a preempted VM resumes where it stopped. Each run records per-step timings, peak memory and errors per series, which util/executionAnalytics uses to measure what a run cost and to predict the cost of the next one.

Running on Terra

You need a Terra account and a Google Cloud billing project. If Terra or Dockstore is new to you, start with Terra's guides and Dockstore's.

1. Open SegmentatorTwoVmWorkflowOnTerra on Dockstore, choose Launch with → Terra, and pick a workspace.
2. Under Data, upload a data table with one row per batch and a SeriesInstanceUIDs column listing the IDC series in that batch. make_terra_manifest.py builds one from an IDC cohort; here is a sample.
3. Under Workflows, open the workflow, choose Run workflow(s) with inputs defined by data table, and select the table and the rows to run.
4. Under Inputs, upload the preset for a model (MOOSE, TotalSegmentator, TotalSegmentator lung vessels) and set yamlListOfSeriesInstanceUIDs to this.SeriesInstanceUIDs. Under Outputs, choose Use defaults, so the outputs are linked from the data table.
5. Click Run analysis.

To submit and monitor many batches from the command line, use util/terraOps, described in operations.md.

Repository contents

Path	Contents
workflows/harmonized/	The harmonized workflow's WDL, and its docs: overview and inputs, operations, development
workflows/common/	Its shared notebooks (convert, output conversion), the base and output-conversion Dockerfiles, and the codes for radiomics features
workflows/models/	Per model: the inference notebook, Dockerfile, input presets and SNOMED overrides
workflows/TotalSegmentator/, workflows/MOOSE/	The older per-model workflows: WDL, notebooks, Dockerfiles, cohort queries
util/terraOps/	Command-line tools to upload data tables, submit batches, watch them and check their outputs
util/executionAnalytics/	Measure what a Terra run cost; fit a cost model on a pilot and predict a larger run
util/pricingOptimization/, util/downloadBenchmarking/, util/Monitoring/	One-off notebooks and queries: GPU prices by region, download speed, GPU quota use
paper/	Scripts and data behind the figures of the 2026 paper; every figure can be redrawn without credentials
.github/workflows/	CI that builds the Docker images and pushes them to Docker Hub (imagingdatacommons/cloudsegmentator-* for the harmonized workflow)
AGENTS.md, .claude/skills/	Notes and procedures for coding agents

Citation

Thiriveedhi VK, Krishnaswamy D, Clunie D, Pieper S, Kikinis R, Fedorov A. Cloud-based large-scale curation of medical imaging data using AI segmentation. Research Square (2024). https://doi.org/10.21203/rs.3.rs-4351526/v1

Release v1.2.0 is archived on Zenodo: https://doi.org/10.5281/zenodo.10712896. A paper covering both runs is in preparation.

License

Apache 2.0; see LICENSE.