Use this static guide to understand local versus cloud mode, serial versus parallel execution, case selection, segmentation choices, rerun rules, reports, and common failures.
For most users, the simplest workflow is best. Choose the dataset, keep the default settings, run the full pipeline, and review the generated previews and reports.
The dataset name organizes outputs. The CT folder should point to the selected CT dataset or its parent folder. Local CT files commonly end in _0000.nrrd.
Run All Stages performs segmentation and downstream phenotype extraction. If you already have masks, switch Segmentation Input to Supply My Own Segmentations.
A single de-identified case is the safest way to verify paths, segmentation quality, report rendering, and runtime before starting a larger batch.
The final labeled visualization, segmentation preview, and case report help confirm that the numbers are connected to plausible anatomy.
The same overall workflow can run on the local machine or through the hosted cloud workflow. The right choice depends on data handling, compute needs, and cost tolerance.
aorta-health, region us-central1, bucket aorta_health_storage, web service aorta-health-web, and worker job aorta-health-worker. The app knows the configured region, but it does not expose the exact physical Google data center behind that region.
Serial and parallel are cloud execution choices. They are different from the local Per Case versus Per Stage processing mode.
Runs one case at a time. This is the safer default for early testing because it is easier to debug and easier to control cost.
Runs multiple cases at once on separate cloud workers. It can finish a batch sooner, but uses more simultaneous CPU and memory and can hit quota or budget limits.
Start with Serial or parallelism 2. Increase only after a small test run succeeds and the team is comfortable with runtime, logs, and cost behavior.
In cloud Serial mode, keeping one worker alive can make repeated jobs feel closer to local development. It avoids repeated warmup delay, but the worker keeps billing until End Session or the inactivity timeout stops it.
| Setting | What it changes | Best use |
|---|---|---|
| Serial | One cloud worker processes one case at a time. | Cost-safe testing, debugging, first run on a dataset. |
| Parallel | Multiple cloud workers process cases at the same time. | Larger batches where wall-clock time matters and cost/quota are acceptable. |
| Per Case | Completes all selected stages for one case before moving to the next. | Review finished case outputs as soon as possible. |
| Per Stage | Runs one selected stage across all cases, then moves to the next stage. | Refreshing one stage across a dataset or debugging a stage-specific issue. |
The stage labels are visible in the app. You do not need to understand the internal notebooks to use Aorta Health, but the stage sequence explains why earlier outputs matter.
Creates an anatomy mask using nnU-Net or prepares supplied masks for downstream analysis.
Checks the basic orientation and prepares the anatomy for spatial reasoning.
Finds the aortic backbone and the aortic bifurcation, using redundant signals so the downstream analysis has a stable landmark.
Adds anatomical context so the pipeline can reason about front-back relationships instead of treating the aorta as a simple tube.
Identifies branch candidates and named branch relationships that help make neck, renal, and abdominal measurements more interpretable.
Builds centerline and cross-sectional geometry used for diameter, area, tortuosity, and quality-control views.
Extracts structured quantitative outputs such as maximum transverse diameter, aneurysm volume, thrombus burden, calcium burden, tortuosity, and related morphology.
Creates the labeled visual review artifacts that connect the measurements back to anatomy.
Advanced Configuration changes what runs, which cases are included, and whether old outputs are reused.
Inference Segmentations means Aorta Health creates masks. Supply My Own Segmentations means phenotype extraction uses masks you already have.
Fast is the usual first choice. High Quality is slower and more conservative, useful for smaller batches or cases needing closer review.
All Cases runs every matched case, subject to filters. One Case is best for testing, troubleshooting, or inspecting a specific case.
Use a case number range, name pattern, or both. Example: range 2-50 plus pattern ^AVT includes AVT_02 and excludes SBUH_10.
Run alphanumerically, by segmentation size, or by oldest selected output. Sorting by last completed output is useful when refreshing stale or missing work.
Always Run regenerates outputs. Only If Missing saves time. Notebook Changed reruns updated analysis logic. Older Than Threshold refreshes stale outputs.
Aorta Health writes structured outputs so users can review individual cases and whole datasets without manually tracking every file.
The full report summarizes the latest known result for the dataset. It may include outputs generated by different runs if some cases were refreshed later.
The case report focuses on one case and is usually the best place to inspect case-level outputs and links.
The metrics report gives table-style quantitative outputs, useful for screening plausibility and comparing cases.
Interactive visualizations help identify whether segmentation, branch labels, centerline geometry, and final measurements look reasonable.
The chatbot is helpful for workflow questions and troubleshooting, but it is separate from CT processing.
Open the relevant item below. Most failures come from path mismatch, missing segmentation, reused stale output, or an interrupted run.
This usually means phenotype extraction was started for a case that does not have a paired segmentation mask.
CASE_0000.nrrd.