DicomRTTool

Published! See the Technical Note and please cite it if you find this work useful. DOI: https://doi.org/10.1016/j.prro.2021.02.003

Convert DICOM images and RT structures into NIfTI files, NumPy arrays, and SimpleITK image handles — and convert prediction masks back into RT structures.

Intended use

DicomRTTool is a reference implementation for research and for building clinical workflows — not a validated clinical device. It has no regulatory clearance of any kind, and nothing in this repository removes an adopting site's responsibility to commission it locally.

What it is for: converting DICOM image series and RT structure sets into arrays, NIfTI, and SimpleITK handles for research and model development; surveying and indexing DICOM corpora; resampling; anonymized export; and writing segmentation output back into an RT structure set.

What it is not for: use as the sole check on any structure set that will be treated from. Two paths in this library produce or modify DICOM that can be imported into a treatment planning system:

• prediction_array_to_RT() writes a model's segmentation into a new RT structure set. Contours produced this way are a starting point requiring review and editing by a qualified person before any clinical use. A rasterization or coordinate error is not always visually obvious.
• rewrite_RT() renames ROIs in an existing structure set and saves the file in place. A wrong association silently relabels a structure, and the in-place write leaves no original to compare against. Keep an independent copy of any structure set before rewriting it.

If output from this library will inform a clinical decision, commission it first: see docs/COMMISSIONING.md and run the validation package described there.

Versioned use

Clinical or clinically-adjacent use is supported only from tagged releases — never from main. main receives development work that has not been through a release check, and its behaviour can change without notice.

Pin an exact version, and record the version you commissioned:

pip install DicomRTTool==6.2.0

Re-commission after any version change that touches conversion, geometry, or the RT write paths; the CHANGELOG flags those.

Installation

pip install DicomRTTool

For the interactive image viewer (requires matplotlib):

pip install "DicomRTTool[viewer]"

Supported Python versions: 3.10, 3.11, 3.12, 3.13.

Getting started: a typical workflow

A first pass through a new DICOM corpus usually moves from discover → survey → select → export:

1. Discover — walk the folder tree and list the ROIs that are present.
2. Survey — write a metadata manifest of everything found (spacing + ROI volumes).
3. Select — choose the ROIs you want and map their name aliases.
4. Export — write NIfTI files, resampled to your target voxel spacing.

Everything else in this README (loading a single series into NumPy, writing predictions back to RT structures, anonymization, performance tuning, …) builds on these four steps and is covered afterwards.

Step 1 — Discover: walk the folders

walk_through_folders recursively scans a directory tree, groups files by SeriesInstanceUID, and links each RT structure and dose to its image series. The images and RT files do not need to live in the same folder.

from DicomRTTool.ReaderWriter import DicomReaderWriter

reader = DicomReaderWriter()
reader.walk_through_folders("/path/to/dicom")

# What ROIs exist across everything that was found?
all_rois = reader.return_rois(print_rois=True)

Step 2 — Survey: write a metadata manifest

Before committing to an export, get a one-row-per-series overview with create_manifest. With no ROIs selected yet it records every ROI it found, so you can see what is available and how large each structure is:

reader.create_manifest("/path/to/manifest.csv")

Each row has patient_hash / study_hash / series_hash, the image spacing (spacing_x/y/z), and one <roi> cc column per ROI giving its mask volume in cubic centimetres (blank when that ROI is absent from the series). Add anonymize=True, salt="MyProjectSalt" to put the deterministic hashes in those columns; otherwise they hold the original PatientID / StudyInstanceUID / SeriesInstanceUID. An anonymization_key.json reverse-lookup file is written next to the manifest so you can review the table and its key together. Re-runs update the file in place (see Incremental manifests), so you can keep growing one manifest as you walk more data.

Step 3 — Select: choose ROIs and map aliases

Real-world ROI names are inconsistent (Lung_L, Lung-Left, left lung). ROIAssociationClass maps any number of aliases onto one canonical name, and set_contour_names_and_associations picks the ROIs you actually want:

from DicomRTTool.ReaderWriter import ROIAssociationClass

reader.set_contour_names_and_associations(
    contour_names=["lung_l", "lung_r", "cord"],
    associations=[
        ROIAssociationClass("lung_l", ["lung-left", "left lung"]),
        ROIAssociationClass("lung_r", ["lung-right", "right lung"]),
        ROIAssociationClass("cord", ["spinal cord", "spinalcord"]),
    ],
)

# Which series contain the selected ROIs?
print(reader.indexes_with_contours)

ROI names are matched case-insensitively. Build the reader with require_all_contours=False to also include series that carry only some of the selected ROIs.

Step 4 — Export: NIfTI with voxel resampling

write_to_folder writes every selected series to a tidy per-case tree — one file per ROI — plus a manifest.csv. Pass output_spacing (mm) to resample on the way out: linear interpolation for the image and dose, nearest-neighbour for masks (labels are never blended). The dose is resampled onto the resampled image grid, so the image, masks, and dose all come out the same size and geometry.

# Build with get_dose_output=True if you also want the dose loaded/resampled.
reader.write_to_folder(
    "/path/to/out",
    output_spacing=(1.0, 1.0, 3.0),     # omit to keep native spacing
    anonymize=True, salt="MyProjectSalt",
)

out/
  <patient>/<study>/<series>/   # hashes when anonymized, else sanitised IDs
    image.nii.gz
    masks/
      lung_l.nii.gz
      lung_r.nii.gz
      cord.nii.gz
    doses/                      # only when get_dose_output=True and dose exists
      plan.nii.gz
    metadata.json               # per-series DICOM features (grouped, schema v2)
  manifest.csv                  # identifiers, spacing, per-ROI volume (cc)
  anonymization_key.json        # only when anonymize=True (reverse lookup)

This nested patient/study/series layout mirrors the companion C# DICOM→NIfTI tool. The manifest.csv has the same shape as the one from Step 2. The metadata.json groups the image / structure / dose / plan features per series (see Grouped metadata; pass metadata_style="flat" for the historical requested-tags-only dict).

The method name implies the breadth: skip ROI selection entirely (no Contour_Names, no rois=) and write_to_folder exports every image series as image + dose only — handy when you just want the images. You can also set anonymize (and salt) once at construction to make it the default for every export, and still override it per call:

reader = DicomReaderWriter(anonymize=True, salt="MyProjectSalt")   # default on
reader.walk_through_folders("/path/to/dicom")
reader.write_to_folder("/path/to/anon")                  # uses the default
reader.write_to_folder("/path/to/clear", anonymize=False)  # override off

That's the core loop, and Examples/01_DICOM_to_NIfTI_Dataset.ipynb runs it end to end on a public TCIA cohort — CT, structures, and dose — through to a verified, anonymized NIfTI dataset. See Examples/.

The sections below are reference material for everything else.

Load a single series into NumPy / SimpleITK

For in-memory analysis (e.g. feeding a model) instead of exporting files, load one series directly:

reader.set_index(reader.indexes_with_contours[0])
reader.get_images_and_mask()

image_numpy  = reader.ArrayDicom         # NumPy image array
mask_numpy   = reader.mask               # NumPy mask array
image_handle = reader.dicom_handle       # SimpleITK Image
mask_handle  = reader.annotation_handle  # SimpleITK Image

Anonymized export

anonymize=True (on write_to_folder or create_manifest) replaces identifiers with deterministic SHA-256 hashes (patient MRN → patient hash, study hash, series hash). For write_to_folder the case folder is named by the series hash and an anonymization_key.json reverse-lookup file is written alongside the manifest. The hashing matches the companion C# tool byte-for-byte, so both tools produce identical hashes for the same salt:

reader.write_to_folder("/path/to/out", anonymize=True, salt="MyProjectSalt")

# Stand-alone helpers are exported too:
from DicomRTTool import hash_patient, hash_study, hash_series, AnonymizationKey
hash_patient("1234567")          # -> 'P...'  (prefix + 5 bytes of SHA-256)

Metadata manifest details

create_manifest (Step 2) defaults to every ROI discovered during the walk. To restrict it to specific ROIs, either set Contour_Names on the reader (as in Step 3) or pass them explicitly:

reader.create_manifest("/path/to/manifest.csv", rois=["tumor", "cord"])

Image rows and dose rows

The manifest holds two kinds of row, told apart by the modality column:

A dose row is keyed by the dose series' own patient_hash / study_hash / series_hash, so it shares the patient and study of the images it belongs to but carries its own series identifier. Its spacing_x/y/z is the dose grid's own spacing, which is usually finer than the image's. Each row only ever fills its own kind of column, so the other kind is blank.

The plan name is taken from the referenced RT Plan's RTPlanName, falling back to its RTPlanLabel. Many public collections ship dose without the plan, in which case the dose's own series description is used instead.

Dmax is read off the native dose grid, before any output_spacing resampling — interpolating a dose onto a coarser grid blunts its peak, and the manifest should report the real maximum. DoseUnits of GY and CGY are both converted to cGy; anything without an absolute equivalent (e.g. RELATIVE) leaves the cell blank and logs a warning.

create_manifest reports every dose the walk found. write_to_folder's manifest describes what it actually wrote, so it includes dose rows only when the reader was built with get_dose_output=True.

Incremental manifests

If the target CSV (and its anonymization_key.json) already exist, create_manifest reads them and updates in place instead of overwriting. Rows for series in the current walk are recomputed and upserted — matched on the series_hash column, so an existing series is updated and a new one is appended — while series not in the current walk are left untouched. New ROI columns are added (left blank for the rows that predate them), and the existing key file's hash mappings (and salt) are reused so identifiers stay stable. This makes it safe to call repeatedly as you walk more data:

# First batch
reader.walk_through_folders("/data/batch1")
reader.create_manifest("/path/to/manifest.csv")

# Later: walk more patients and keep populating the same file —
# existing rows are preserved, only new series are added.
reader.reset()
reader.walk_through_folders("/data/batch2")
reader.create_manifest("/path/to/manifest.csv")

(write_to_folder writes the same-shape manifest alongside the NIfTI tree; use create_manifest when you want the table on its own or want to grow it over multiple runs.)

Resampling any SimpleITK handle

The resampling helpers used by the writers are exported for direct use:

from DicomRTTool import resample_to_spacing, resample_to_reference

# To a target voxel spacing (linear for images/dose, "Nearest" for masks):
resampled = resample_to_spacing(reader.dicom_handle, (1.0, 1.0, 3.0), "Linear")

# Onto another image's exact grid (size/spacing/origin/direction):
dose_on_image = resample_to_reference(reader.dose_handle, resampled, "Linear")

write_images_annotations also accepts output_spacing for the single-series combined-file output (Overall_Data_* / Overall_mask_* / Overall_dose_*).

Writing predictions back to an RT structure

import numpy as np

# 4-channel one-hot prediction matching the loaded image shape:
# (slices, rows, cols, num_classes + 1) — channel 0 is background.
predictions = np.zeros((*reader.ArrayDicom.shape, 3), dtype=np.float32)
# ... populate `predictions` from your model ...

reader.prediction_array_to_RT(
    prediction_array=predictions,
    output_dir="/path/to/output",
    ROI_Names=["organ_a", "organ_b"],
)

Reading extra DICOM tags

Pull additional tags by name. SITK keys (image_sitk_string_keys, dose_sitk_string_keys) use "group|element" strings; pydicom keys (plan_pydicom_string_keys, struct_pydicom_string_keys) use Tag objects:

from pydicom.tag import Tag

reader = DicomReaderWriter(
    image_sitk_string_keys={"MyPatientName": "0010|0010", "Manufacturer": "0008|0070"},
    plan_pydicom_string_keys={"MyNamedRTPlan": Tag((0x300a, 0x0002))},
)
reader.walk_through_folders("/path/to/dicom")

# Per series, the pulled values are on the entry:
entry = reader.series_instances_dictionary[0]
print(entry.additional_tags)        # {"MyPatientName": ..., "Manufacturer": ...}

When you export with write_to_folder, these requested tags are also written into each series' metadata.json — inside the owning category's tags sub-dict of the grouped document described below (or as a bare {name: value} dict with metadata_style="flat"). Note that the values are written verbatim — if you anonymize the folder names, make sure the tags you pull don't themselves carry identifying information.

Grouped metadata (schema v2)

By default write_to_folder writes a versioned metadata.json for every exported series (metadata_style="grouped"). DICOM features the walk already parsed are grouped by category — image (modality, series description, frame of reference, pixel spacing / slice thickness, effective export spacing), structures (each ROI's name, number, interpreted type, structure code, plus volume_cc and exported_file for exported ROIs), doses (dose units / type / summation type, referenced plan & struct UIDs, included_in_sum, plus the plan_name and dose_max_cgy that identify the dose in the manifest), dose_file, and plans (label / name). Your *_string_keys requests land in each category's own tags sub-dict. Categories with no corresponding DICOM files are simply omitted, so image-only or dose-less folders parse cleanly with meta.get("doses", []). The PHI note above applies unchanged: requested tag values are written verbatim.

Pass metadata_style="flat" for the historical behavior — a flat {name: value} dict of the requested tags, written only when some were found.

Resetting state between uses

DicomReaderWriter instances can be reused across multiple corpora; call the appropriate reset method before walking a fresh folder tree or swapping target ROIs:

reader.reset()        # wipe everything (images, RTs, masks, cached UIDs)
reader.reset_rts()    # clear ROI bookkeeping only; keep loaded images
reader.reset_mask()   # re-allocate an empty mask after changing Contour_Names

Performance

Both create_manifest and write_to_folder parallelise across series, and auto-tune per-ROI rasterisation threads so a single series with many ROIs still uses your spare cores. You can also set it explicitly — useful when calling get_images_and_mask() directly on one big multi-ROI series:

reader = DicomReaderWriter(Contour_Names=[...], mask_thread_count=4)

mask_thread_count=1 (the default) is the serial path; the parallel path produces byte-identical masks.

Cross-tool evaluation

The evaluation/ directory contains an opt-in harness that runs DicomRTTool and the companion C# DicomRtNifti.Cli tool side by side on TCIA LCTSC patients and compares mask generation (Dice / volume), image generation (voxel MAE / geometry), and the resampling feature. See evaluation/README.md. It is never part of the hermetic test suite — the parity pytest auto-skips unless you point it at the external dataset and the built C# binary.

Reproducibility

Dependency bounds. Every runtime dependency in pyproject.toml carries a lower bound and no upper cap. Lower bounds are the oldest release that ships wheels for our minimum Python (3.10) and provides the API we call; caps are omitted deliberately, because a cap in a library's metadata becomes an unsatisfiable constraint in a downstream environment. The bounds are executable, not decorative: the minimum-versions CI job derives the floors from the installed package metadata, force-installs exactly those versions on Python 3.10, and runs the suite — so a floor that is too low fails CI instead of an installing user.

Pinned conformance suite. The accuracy gate depends on RTMaskConformanceTest, which is pinned to a full commit SHA in requirements-conformance.txt. It is a gate, so it must not be able to change verdict without a reviewed commit here. Upstream publishes no tags or PyPI release, so a SHA is the only immutable reference available; bump it in its own commit.

Recorded environments. constraints/ci-linux-py3.12.txt pins the full transitive closure of the canonical CI environment. It is a constraints file, not a lockfile — nothing in it reaches an installing user, and the floating test matrix ignores it so upstream regressions are still caught. The pinned job installs against it, and the conformance job runs under it, so the record is proven to still resolve and still pass. Every CI job additionally uploads its own resolved pip freeze as an artifact (including on failure), and the publish workflow uploads build provenance — the build environment, interpreter, commit, and SHA-256 of each distribution — alongside the dist itself.

What's new since v4.0

• write_to_folder — bulk DICOM→NIfTI export to a per-ROI layout (<case>/image.nii.gz, <case>/masks/<roi>.nii.gz, <case>/doses/…) with a single manifest.csv and no persistent index.
• create_manifest — write (or incrementally extend) a metadata-only CSV of per-series image spacing and per-ROI volumes, mirroring the C# manifest.
• Output resampling — output_spacing on write_to_folder and write_images_annotations, plus the public resample_to_spacing / resample_to_reference helpers (linear for image/dose, nearest-neighbour for masks; dose lands on the resampled image grid).
• Anonymization — deterministic SHA-256 hashing (hash_patient / hash_study / hash_series / AnonymizationKey) for anonymized exports, matching the companion C# tool.
• Faster, parallel rasterisation — mask_thread_count plus the removal of a per-ROI full-array rescan (~2.4× on multi-ROI series).
• Cross-tool evaluation harness under evaluation/.

What's new in v4.0

• Python 3.10+ required (3.8 / 3.9 are end-of-life).
• Public state-reset API: reset(), reset_rts(), reset_mask() — replaces the v3 __reset__ / __reset_mask__ / __reset_RTs__ accessors.
• Deprecated v3 names removed: down_folder → walk_through_folders, where_are_RTs → where_is_ROI, with_annotations → prediction_array_to_RT, plus the __set_iteration__ and __set_description__ setters renamed to set_iteration / set_description. See CHANGELOG.md for the full list and migration notes.
• Excel → CSV for both bulk-export helpers, dropping the openpyxl dependency: characterize_data_to_excel is now characterize_data_to_csv, and write_parallel(excel_file=…) is now write_parallel(index_file=…) accepting a .csv path.
• struct_pydicom_string_keys plumbing finally works — historically the parameter was accepted but the values never reached the parsed RT records.
• Architecture: the original ReaderWriter.py god-class has been partly extracted into a new internal _internal/ package. The public DicomReaderWriter API is unchanged.
• Hermetic test suite: every DICOM file the tests need is generated in a tmp directory at session start from analytical primitives. No external corpus, no network, no caches — the full suite runs in ~6 seconds and validates against analytically-known volume truth.
• Tooling: ruff replaces flake8; PyPI Trusted Publishing replaces the PYPI_TOKEN secret; CI matrix expanded to ubuntu + windows × four Python versions; pre-commit config added.

License

GPL-3.0-or-later

Citation

If you find this code useful, please cite both:

• The technical note — Anderson, Wahid, and Brock, Practical Radiation Oncology 11(3):226–229, 2021. doi:10.1016/j.prro.2021.02.003
• The software — doi:10.5281/zenodo.21631862. This is the concept DOI and always resolves to the latest release; Zenodo also mints a version-specific DOI per release if you need to pin the exact version you ran.

CITATION.cff carries both, so GitHub's "Cite this repository" button and most reference managers pick them up without any manual entry.