Medical imaging

DICOM packaging, retained and removed tags, report format, pseudonymous IDs and the study manifest.

Updated 15 Oct 2026 · v1

Radiology is licensed as studies, not loose images: full DICOM series with the radiology report, and EMR where the source can link it. Everything is de-identified before it leaves the site.

Packaging

One folder per study, a subfolder per series, and the report beside the images.

QD-9M4R / a CT datasetlayout
  • QD-9M4R/
  • manifest.jsonlone record per study
  • datasheet.pdf
  • deid-report.pdf
  • studies/
  • st_7Q2KX9M1/pseudonymous study ID
  • study.jsonthe study manifest
  • report.txtredacted radiology report
  • emr.jsonwhere the site links it
  • series/
  • se_01/axial, soft tissue
  • IM_0001.dcm … IM_0412.dcm
  • se_02/axial, lung

Tags

DICOM headers are de-identified to the PS3.15 Basic Application Level Confidentiality Profile, with the Clean Pixel Data, Clean Descriptors, Retain Longitudinal Temporal Information with Modified Dates, and Retain Patient Characteristics options.

KeptRemoved or replaced
Modality, Body Part Examined, Study and Series Description (cleaned)Patient Name, Other Patient IDs, Patient Address, phone numbers
Manufacturer, Manufacturer Model NameDevice serial numbers, station names
Acquisition parameters: Slice Thickness, KVP, Magnetic Field Strength, Repetition and Echo Time, Contrast agentInstitution Name and Address (region of site kept in metadata)
Patient Sex; age as a bandExact birth date; Accession Number; physician and operator names
Pixel spacing and geometryBurned-in text in pixel data (detected and masked)
Study, Series and SOP Instance UIDs, remapped consistentlyOriginal UIDs

Pseudonymous IDs

Each patient and study gets a stable pseudonym (pt_…, st_…) derived inside the owner's environment with a keyed hash. The key never leaves the site. The same patient keeps the same pseudonym across a dataset, so prior and follow-up studies stay linked; the pseudonym can't be reversed outside the owner's environment. Dates are shifted by a consistent offset per patient, so intervals between studies are preserved.

Reports

report.txt is the radiology report in UTF-8 plain text, with free-text redaction applied (names, dates, places, IDs replaced by tags such as [DATE]). Findings and impression are kept as written. Mammography reports carry the BI-RADS category, also lifted into study.json.

EMR fields

Where the site can link EMR, emr.json carries diagnoses (with ICD-10 codes where the source has them), relevant lab results, encounter type, follow-up studies, and biopsy outcome for mammography. All dates are shifted with the study's offset.

The study manifest

{
  "study_id": "st_7Q2KX9M1",
  "patient_id": "pt_K4D8N2QW",
  "dataset_id": "QD-9M4R",
  "modality": "CT",
  "body_part": "CHEST",
  "manufacturer": "SIEMENS",
  "model": "SOMATOM Force",
  "series": [
    { "series_id": "se_01", "description": "AXIAL SOFT TISSUE", "images": 412, "slice_thickness_mm": 1.0, "kvp": 120, "contrast": true },
    { "series_id": "se_02", "description": "AXIAL LUNG", "images": 412, "slice_thickness_mm": 1.0, "kvp": 120, "contrast": true }
  ],
  "age_band": "50-59",
  "sex": "F",
  "site_region": "IN-South",
  "report": "report.txt",
  "emr": "emr.json",
  "date_offset_applied": true
}

The values above are a specimen. Each dataset's modality mix, manufacturers and counts are on its deal sheet.