CloudSegmentator runs segmentation models (MOOSE, TotalSegmentator) on IDC imaging data in the cloud and converts the results to DICOM-SEG, radiomics and DICOM SR.
workflows/harmonized/— the model-agnostic workflow (one WDL, three notebooks). Start atworkflows/harmonized/Docs/README.md.workflows/common/,workflows/models/<model>/— its shared and per-model notebooks, Dockerfiles and resources.workflows/MOOSE/,workflows/TotalSegmentator/— the older per-model workflows; keep changes to them separate.util/terraOps/— command-line tools to submit and monitor Terra runs.util/executionAnalytics/— cost measurement and prediction.
| Task | Read |
|---|---|
| Changing notebooks or the WDL, reproducing a failure | workflows/harmonized/Docs/development.md |
| Submitting, monitoring or triaging Terra runs; checking delivered DICOM | workflows/harmonized/Docs/operations.md |
| Measuring or predicting cost | util/executionAnalytics/README.md |
Claude Code users get these as project skills in .claude/skills/
(pipeline-dev, terra-runs, dicom-delivery, cost-analysis).
- Never guess cloud targets. Terra workspace, method config, image registry,
and delivery bucket come from flags or
TERRA_WORKSPACE,TERRA_METHOD_CONFIG,SEGMENTATOR_REGISTRY,SEGMENTATOR_DELIVERY_BUCKET. If one is unset, ask. - Confirm before spending or publishing: submitting Terra jobs, aborting them, pushing Docker images, and uploading to buckets or DICOM stores all need the user's go-ahead.
- Notebooks: edit cell sources surgically, never re-serialize a whole notebook, never commit outputs (see development.md).
- No run data in git: per-submission CSVs, billing exports, manifests and
figures from individual runs stay local (
.gitignorecovers the common ones).