Skip to content

8.2 Harmonize names

Research datasets often use inconsistent Series Description text. Harmonize suggests a standard name from anatomy and (for CT/MR) contrast — RSNA Radiology Playbook / RadLex style. It does not change pixels.

Demo series

Fixture Path under tests/controller/assets/test_dcm_files/ What it shows
CT_Head_With_Contrast CT_Head_With_Contrast CT path: TotalSegmentator anatomy / contrast, brain-structures prompt, overlays
davidson_cxr davidson_cxr XR planar path: optional Xp-Bodypart + CXp-Projection-Rotation
us_rgb_single_frame us_rgb_single_frame US planar path: DICOM metadata only (no TotalSegmentator)

Import a series, open it in Series View, then run Harmonize. Reuse CT_Head_With_Contrast in 8.3 Blur faces.

Goal

  1. On CT_Head_With_Contrast: run Harmonize Description, answer the brain-structures prompt when offered, Apply, then review segmentation overlays on the middle slice.
  2. On davidson_cxr and us_rgb_single_frame: run Harmonize and read the Playbook table — Source names the model or DICOM (same idea as CT’s “TotalSegmentator anatomy”).

Before you start

  1. Complete AI Features setup — Harmonize CT pack, resolution, and Brain structures for the CT demo.
  2. Optionally download XR body part and XR chest view so the CXR demo shows pixel fusion (without them, XR still Harmonizes from DICOM tags).
  3. Import the demo series (Search) and open each in Series View (View — right-click the series row).

Workflow on CT_Head_With_Contrast

1. Harmonize Description (from Series View)

  1. With CT_Head_With_Contrast open in Series View, click Harmonize Description.
  2. Wait until analysis finishes — the Playbook harmonization table fills with anatomy / contrast evidence (not an empty table mid-progress).
  3. Review the suggested Series Description, then Yes to apply, or No / Cancel.

The Series View remains behind the dialog so you keep context for the open series.

Series View with completed Harmonize Description results

2. Brain structures prompt

On a CT head series, Harmonize asks whether to run detailed brain structure segmentation before the job continues:

  1. Read the Brain structures Yes/No message (academic license / models required — see AI Features).
  2. Choose Yes to include brain structures in this run, or No for standard anatomy only.

Harmonize Description with brain structures prompt

3. Segmented Series View

After you Accept a Harmonize run that included brain structures (Yes on the prompt):

  1. Series View shows latch buttons for whole brain plus detailed structures (brainstem, lobes, ventricles, …) when the licensed pack ran.
  2. Select all structures you want to review (including brain).
  3. Go to the middle slice to see overlays on a representative frame.

Series View with all brain segments on the middle slice

Planar Harmonize (XR and US)

CR/DX, ultrasound, and mammography use a separate Harmonize path from CT/MR: no TotalSegmentator, no thickness/contrast buckets. The Playbook table still uses Evidence (what was measured) and Source (where it came from).

Chest X-ray (davidson_cxr)

  1. Open davidson_cxr in Series View → Harmonize Description.
  2. When XR models are installed, Body Part Source is Xp-Bodypart (or fused with DICOM); View / Rotation use CXp-Projection-Rotation when anatomy is Chest.
  3. Evidence looks like CT classifier rows — e.g. Chest · 99.00% confidence — not opaque pixel:… tokens.
  4. Rotation is analysis-only; SeriesDescription stays e.g. Chest AP / Chest Lat.

davidson_cxr Series View above planar Harmonize with XR model sources

Ultrasound (us_rgb_single_frame)

  1. Open us_rgb_single_frame in Series View → Harmonize Description.
  2. Cohort / body part / mode come from DICOM tags and keywords; Source is DICOM metadata or RadLex Playbook.
  3. No XR pixel packs and no TotalSegmentator run on US.

us_rgb_single_frame Series View above planar Harmonize (metadata-only)

Notes (same as production use)

  • Scouts, MIP/VR, dose reports, and similar series are usually skipped.
  • CT: anatomy segmentation + contrast phase when models allow (TotalSegmentator).
  • MR: anatomy from MR packs; IV contrast from DICOM headers (TotalSegmentator).
  • XR (CR/DX): planar Harmonize; optional Xp-Bodypart + chest-only CXp view/rotation (soft downloads).
  • US / MG: planar Harmonize from DICOM only — no TotalSegmentator and no XR pixel packs.
  • SC / OT / DOC: Harmonize is not offered.
  • After all series in a study are harmonized, the best LOINC study description is applied automatically (same path as AI Batch). Change it later from Dataset. Pure XR/US/MG studies use the matching LOINC prefix.
  • In Dataset, double-click a study or series description (or multi-select and right-click for Set description) to pick LOINC (study) or RadLex (series) names — including for rows not yet green. Single-click only selects the row.
  • Results cache under the series folder — Clear Analysis Cache for a fresh CT/MR run.
  • Batch: 8.4 Run on many studies (resolution comes from AI Features, not per batch).

What good looks like

  • CT: SeriesDescription looks consistent for this head CT; Dataset Harmonized updates; after brain Yes, latch overlays draw on the middle slice.
  • CXR: Playbook Source names Xp-Bodypart / CXp-Projection-Rotation (or DICOM) with readable confidence evidence.
  • US: Playbook rows cite DICOM metadata / RadLex Playbook; suggested name matches ultrasound anatomy/mode.

If it fails

  • Not suitable series → expected skip.
  • Already harmonized → Clear Analysis Cache to re-run.
  • Models not ready → AI Features.

Next steps

Continue with 8.3 Blur faces on the same CT_Head_With_Contrast series (Gaussian).