dicom-forge

A medical-imaging research pipeline that prepares DICOM for 3D Slicer and ITK-SNAP.

dicom-forge takes a folder of DICOM, refuses series whose metadata signals burned-in pixel text, de-identifies the headers (names, IDs, private tags, UIDs), runs quality control, and converts it into formats clinical-research viewers load natively — NIfTI (.nii.gz), NRRD (.nrrd), and Slicer's segmentation format (.seg.nrrd). Every run produces a serialisable audit record.

It is the headless core of a two-repo system; its companion, slicer-forge, wraps it in a 3D Slicer extension GUI.

📌 New here? Read the project showcase for the big picture — the problem, the architecture, who it's for, and the engineering highlights.

---

Why this exists

Real imaging pipelines separate a testable, headless core from a thin GUI. 3D Slicer itself is built this way: ITK/VTK do the work, the GUI is a shell on top. dicom-forge is that core — so the logic is unit-tested in CI without needing Slicer installed, and the same code runs from the command line, from Python, or inside Slicer's Python console.

Features

• Ingestion — recursive DICOM discovery, series grouping, geometry-aware slice ordering (handles oblique acquisitions), rescale-slope/intercept applied (CT in HU).
• Header de-identification — three levels aligned with a documented subset of the DICOM PS3.15 Basic Profile, applied recursively. It removes names, IDs and private tags, replaces UIDs consistently, and screens free text. PatientID and UIDs use keyed HMAC pseudonyms under a private project secret. Series whose metadata signals burned-in pixel text are rejected by default.
• Quality control — slice count, geometry consistency, slice-spacing regularity, and intensity statistics, split into blocking errors vs. non-blocking warnings.
• Conversion — geometry-preserving export to NIfTI / NRRD via SimpleITK (the ITK core shared with Slicer & ITK-SNAP).
• Segmentation — write Slicer-native .seg.nrrd label maps with named, coloured segments.
• Typed & validated — Pydantic models everywhere, py.typed, strict mypy.
• Two interfaces — a Rich CLI (dicomforge) and a clean Python API.
• Native ITK pre-flight — a companion C++/ITK CLI, dicom-probe, reads a series through ITK's own GDCM path (the engine under Slicer/ITK-SNAP) and reports the true volume geometry as JSON.

Installation

pip install dicom-anvil            # core (ingest, de-id, QC)
pip install "dicom-anvil[convert]" # + SimpleITK/pynrrd/nibabel for conversion

Names: the package installs as dicom-anvil (the PyPI name dicom-forge was already taken) and imports as dicomforge; the repository stays dicom-forge.

The conversion stack (SimpleITK et al.) is an optional extra so the core stays lightweight and CI-friendly. Calling a conversion function without it raises a clear MissingDependencyError telling you exactly what to install.

Quick start

Command line

# List the series in a folder
dicomforge inspect ./study

# Run QC and print a report
dicomforge qc ./study

# A private per-project secret keeps pseudonyms stable across runs (never commit it)
export DICOMFORGE_PSEUDONYM_SECRET="<from your secrets manager>"

# Full pipeline: burned-in check -> de-identify -> QC -> convert to Slicer NRRD
dicomforge convert ./study ./out/case01 --format nrrd --deid-level moderate

Python

from dicomforge import run_pipeline, PipelineConfig, OutputFormat

result = run_pipeline(
    "./study",
    "./out/case01",
    config=PipelineConfig(output_format=OutputFormat.NRRD),
)
print(result.qc.passed)  # True / False
print(result.pixel_privacy.status)  # "unverified" (never "clean"): pixels are not analysed
print(result.conversion.output_path)  # ./out/case01.nrrd
print(result.model_dump_json(indent=2))  # full audit record

Pipeline order (and why it matters)

ingest ──> burned-in check ──> de-identify ──> QC ──> convert

Series flagged for burned-in pixel annotation are rejected before anything is written. Header de-identification runs in memory before conversion — your source DICOM is never modified. NRRD/NIfTI outputs carry no DICOM header fields, but their voxels are the source pixels, unchanged.

⚠️ Not a validated anonymization tool. De-identification is header-only risk reduction. There is no HIPAA, GDPR or other regulatory claim, and no clinical validation. Pixel data is never analysed: burned-in text is caught only when metadata signals it. Free-text screening finds only the record's own identifiers, and some quasi-identifiers (age, sex; dates at basic) are kept by design. Read the de-identification policy and review output before it leaves a controlled environment.

Development

python -m venv .venv && . .venv/Scripts/activate   # Windows
pip install -e ".[dev,convert]"
pytest                      # run the suite (synthetic DICOM, no real data needed)
ruff check . && mypy        # lint + type-check

License

PolyForm Noncommercial License 1.0.0 © Angela Hudson

Noncommercial use — personal, research, educational, and other noncommercial purposes as defined by the license — is free. Any commercial use requires a separate license from the copyright holder. See LICENSE for the full terms, or open an issue to ask about commercial licensing.