pydcm

The complete DICOM toolkit for Python. A native reader/writer with a built-in decoder for every transfer syntax (no codec plugins), plus zero-copy NumPy / PyTorch pixel access — a compiled extension does the work in-process.

• Decodes JPEG, JPEG-2000, JPEG-LS, RLE, JPEG-XL and HTJ2K out of the box, no plugins, and returns Hounsfield units and 3-D volumes.
• Reads text from any vendor or locale — full Japanese / Korean / Chinese multibyte and every ISO 2022 escape, decoded adaptively to correct UTF-8 where a strict reader would error or garble it.
• One wheel replaces a whole stack of single-purpose tools: read / decode / write every transfer syntax; DIMSE (pydcm.dimse) and DICOMweb (pydcm.dicomweb) networking; NIfTI / NRRD / MetaImage / BIDS / DWI volume export; segmentations, parametric maps and structured reports; PS3.15 de-identification and IOD / SR conformance validation; nnU-Net / MONAI preprocessing (pydcm.transforms, bit-exact); radiomics (pydcm.radiomics); whole-slide imaging (pydcm.wsi); RT dose & DVH (pydcm.rt); waveforms; FHIR / HL7 bridges; and an in-process MCP server (pydcm.mcp) for agents. See the capability map.
• The API is clean and Pythonic, and most existing Python DICOM code runs against it unchanged. SimpleITK images are accepted / returned where natural (radiomics, seg masks) without being a dependency.

Not a medical device. pydcm is not intended or cleared for clinical or diagnostic use. Decoded pixels and HU are for research/engineering only; validate any output for your intended use.

Install

pip install pydcm

A platform wheel ships the compiled extension — no build step, no DICOM library on your side. Wheels are published for macOS (arm64, x86_64) and Linux (aarch64, x86_64; manylinux_2_28, i.e. glibc 2.28+ — RHEL 8+, Ubuntu 20.04+, Debian 11+). One wheel per platform covers CPython 3.12+ (stable ABI).

Use

import pydcm

arr = pydcm.decode("scan.dcm")             # ndarray [frames, rows, cols(, samples)]
frame = pydcm.decode("ct.dcm", frame=2)    # 1-based; 0 (default) = all frames
arr, meta = pydcm.decode("scan.dcm", with_meta=True)

Read & edit

dcmread returns a familiar Dataset. Attribute, item and sequence access, PersonName, file_meta, and a lazy pixel_array all behave the way Python DICOM users expect — so much that import pydcm as pydicom is usually enough:

ds = pydcm.dcmread("scan.dcm")
ds.PatientName              # PersonName — .family_name / .given_name / .ideographic
ds.Rows, ds.PixelSpacing    # 64, [0.3125, 0.3125]   (MultiValue)
ds[0x0010, 0x0010].value    # element access by tag / keyword / (group, element)
ds.file_meta.TransferSyntaxUID
for elem in ds: ...         # iterate in tag order
px = ds.pixel_array         # NumPy (any transfer syntax, no plugins)

Editing preserves everything. save_as patches the original file bytes via the native editor, so the Transfer Syntax, PixelData (including compressed J2K / RLE), private tags and every untouched element survive byte-for-byte:

ds = pydcm.dcmread("ct.dcm")
ds.PatientName = "Anon^Patient"
del ds.PatientBirthDate
ds.save_as("ct_anon.dcm")   # pixels + TS intact; only the named tags change

Verified for near-total element-value fidelity over a large real-world corpus (the lone difference is a private sequence pydcm parses where others leave it opaque — pydcm exposes more). The keyword↔tag↔VR dictionary holds 17,699 entries — a strict superset of the standard dictionaries — so attribute names resolve broadly. Known limits: file_meta surfaces the three mandatory UIDs (not optional group-0002 elements); a from-scratch (no source file) save_as uses a metadata-only path that omits pixels.

HU / real-world values

rescale=False (default) returns the stored integers (lossless). rescale=True returns the modality-LUT output — real-world values, i.e. HU for CT — as float32 (per-frame rescale applied for Enhanced multi-frame):

hu = pydcm.decode("ct.dcm", rescale=True)         # float32 Hounsfield Units
ds = pydcm.DICOMDataset("ct_series/", rescale=True)

Geometry sidecar

with_meta=True also returns the geometry the engine parsed — no extra read:

_, m = pydcm.decode("ct.dcm", with_meta=True)
m["rescale_slope"], m["rescale_intercept"]
m["pixel_spacing"]                  # [row, col] mm
m["image_position_patient"]         # (0020,0032) [x, y, z]
m["image_orientation_patient"]      # (0020,0037) 6-vector
m["slice_thickness"], m["window_center"], m["window_width"]
m["modality"], m["series_instance_uid"], m["sop_instance_uid"]

A directory → PyTorch

DICOMDataset walks a directory and decodes one image per item. It is DataLoader-compatible without importing torch (torch stays optional):

from torch.utils.data import DataLoader

ds = pydcm.DICOMDataset("study_dir/", to_torch=True)   # finds .dcm + extension-less DICOM
for batch in DataLoader(ds, batch_size=8, num_workers=4, shuffle=True):
    ...   # [B, H, W] or [B, H, W, C]

• One sample = one file. Single-frame files yield [H, W] / [H, W, C]; multi-frame files yield [frames, H, W(, C)].
• transform=fn reshapes each sample (e.g. [C, H, W], windowing, scaling).
• pattern="*.dcm" selects by name; the default also detects DICOM by the DICM preamble (catching extension-less clinical exports).

A directory → one 3D volume

load_series assembles a directory of slices into a single spatially-ordered 3D HU volume (IOP clustering + IPP-projection sort, all in the native engine):

vol = pydcm.load_series("ct_series/")
vol.pixels          # ndarray [depth, rows, cols], float32 HU, sorted by position
vol.spacing         # (z, y, x) mm — slice spacing computed from IPP deltas
vol.affine          # 4×4 voxel→world
vol.series_instance_uid

The largest coherent volume in the directory is returned, so a stray localizer or second series does not corrupt the stack.

Beyond the core — capability map

Everything below ships in the same wheel, over the same native engine, with Python kept to thin marshalling — each area verified for correctness against reference data.

Area	Import	What it does
DIMSE networking	pydcm.dimse	SCU + full SCP, persistent associations
DICOMweb client	pydcm.dicomweb	QIDO / WADO / STOW / DELETE, streaming, TS negotiation
DICOM ↔ NIfTI	Volume.to_nifti / from_nifti	gantry-tilt-correct affine, vendor quirks handled; NIfTI → DICOM too
DICOM → NRRD / MetaImage	Volume.to_nrrd / Volume.to_metaimage	3D Slicer (.nrrd) + ITK / nnU-Net (.mha) on-ramps; double-faithful LPS geometry
DWI / diffusion	load_dwi / save_dwi / diffusion_table	FSL .bval/.bvec (feeds FSL / MRtrix / dipy); Siemens CSA + mosaic, enhanced-MF, GE / Philips / UIH private
BIDS sidecar	bids_sidecar	BIDS .json (PhaseEncodingDirection, SliceTiming, EffectiveEchoSpacing…)
Preprocessing transforms	pydcm.transforms	resample / normalize / sliding-window; Tier 1 bit-exact (B-spline convention), Tier 2 ≤ 1 ULP (deep-learning convention)
Whole-slide imaging	pydcm.wsi	tile / region reads on the DICOM WSI pyramid + viewer tiles / total pixel matrix; bit-exact multi-vendor
RT dose & structures	pydcm.rt	read_rtdose / write_rtdose / dvhcalc (full ROI DVH coverage) + write_rtstruct — author RT Structure Sets (ROI contours over a reference series)
Tractography	write_mktract	author Tractography Results (66.6) from track sets — points + per-track measurements / statistics
Radiomics	pydcm.radiomics	the full IBSI set — 135 features / 10 classes; both aggregation conventions; register custom Python features
SEG	write_seg / write_seg_fractional / read_seg	coded SEG, binary + fractional, SEG → labelmap
Parametric Map	write_paramap / read_paramap	author / read float parametric maps
Constructor-style object classes	pydcm.sc / seg / pm / ko / pr / ann / sr	SCImage, Segmentation, ParametricMap, KO, GSPS, the MeasurementReport content-tree classes, MicroscopyBulkSimpleAnnotations
SR / TID 1500	write_sr / write_report / read_report / sr_validate	a full SR object model (Comprehensive / Enhanced / 3D SR, ROI measurement groups, every content-item type) — author from a content tree or as TID 1500 reports, read them back, validate structure + PS3.16 coded concepts, render SR → HTML / JSON
KO / PR / annotations	write_ko / write_pr / read_ann	Key Object Selection, Presentation State, Bulk Annotations
Encapsulated documents	write_encapsulated / read_encapsulated	PDF / CDA / STL / OBJ / MTL (PS3.3 A.104)
Surface meshes	read_surface	read Surface Segmentation (66.5) — points + every mesh primitive (triangle / strip / fan / facet) as (N,3) / (M,3) arrays
Waveforms (ECG / EEG)	pydcm.waveforms	12-lead ECG / EEG read & write; arrays ready for analysis tools (MNE / neurokit2)
FHIR / HL7	pydcm.fhir / pydcm.hl7	DICOM → FHIR R4 ImagingStudy; HL7 v2.x parse (2.3 – 2.8) + ORU^R01 build
De-identification	deident / deidentify / deidentify_series	PS3.15 Table E.1-1 Basic Profile (617 attributes) + every option profile (clean graphics / descriptors / structured content, retain longitudinal / patient-characteristics / device / UIDs / institution); consistent UID remap, dates shifted, CID 7050 method codes stamped in (0012,0064)
Validation	validate / iod_validate / sr_validate	validate is the full report — element-level layers (VR, VM, enumerated values, per-VR format, SpecificCharacterSet, pixel geometry, LUT) plus per-SOP-Class IOD conformance (Type 1/2 mandatory modules, top-level and nested), SR structural / coded / TID content-template checks, and RT cross-reference integrity; iod_validate is the narrower IOD-only view
Rendering / overlays	render_overlay	render a frame to 8-bit RGB and burn GSPS / SR / SEG / RTSTRUCT markup onto it — the frame an agent "sees"
Agent / MCP	pydcm.mcp	in-process MCP server over live pydcm objects
File sets	pydcm.fileset	read a DICOMDIR / File-set, iterate instances

DIMSE networking

import pydcm.dimse as dimse               # drop-in for pynetdicom

ae = dimse.AE(ae_title="PYDCM")
assoc = ae.associate("localhost", 4242, ae_title="ORTHANC")   # dcm4chee: 11112, "DCM4CHEE"
assoc.send_c_echo()
assoc.send_c_store(pydcm.dcmread("ct.dcm"))   # persistent: many ops, one association
assoc.release()

AE.start_server runs the SCP side — EVT_C_STORE / ECHO / FIND / GET / MOVE handlers plus the DIMSE-N set.

DICOMweb client

Address a server by an origin plus a base path — the prefix differs per product (Orthanc /dicom-web, dcm4chee-arc /dcm4chee-arc/aets/DCM4CHEE/rs, a root-mounted server ""):

from pydcm import dicomweb

S, BP = "http://localhost:8042", "/dicom-web"      # Orthanc
studies = dicomweb.search_studies(S, base_path=BP, matches={"PatientID": "42"})
for part10 in dicomweb.iter_study(S, study_uid, base_path=BP):
    ...                                    # streaming retrieve, bounded memory
dicomweb.store_instances(S, [open("ct.dcm", "rb").read()], base_path=BP)

See the DICOMweb and DIMSE how-tos for per-server connection tables, C-FIND/GET/MOVE and TLS.

DICOM ↔ NIfTI, BIDS, DWI

vol = pydcm.load_series("ct_series/")
vol.to_nifti("ct.nii.gz")                  # validated affine (incl. gantry tilt)
vol.to_nrrd("ct.nrrd", gzip=True)          # 3D Slicer (double-faithful LPS space)
vol.to_metaimage("ct.mha", compress=True)  # single-file MetaImage for ITK / nnU-Net
meta = pydcm.bids_sidecar("ep2d_diff/")    # standard BIDS .json fields
dwi = pydcm.load_dwi("ep2d_diff/")         # 4D stack + b-values/b-vectors (all vendors)

Preprocessing transforms

from pydcm import transforms as T

out = T.resample_cubic(vol, out_shape)     # bit-exact B-spline order-3
seg = T.sliding_window_inference(vol.pixels, roi_size=(96, 96, 96), predictor=model)

Tier 1 ops are bit-exact for the classic B-spline convention; Tier 2 ops match the deep-learning (grid-sample) convention to ≤ 1 float32 ULP — same numbers in training and serving, no Python image stack required.

Whole-slide imaging

from pydcm import wsi

slide = wsi.open_slide("wsi_dir/")         # DICOM WSI pyramid (one or many files)
region = slide.read_region((x, y), level=0, size=(512, 512))   # RGBA, level-0 coordinates
slide.associated_images["LABEL"]

RT dosimetry

grid = pydcm.read_rtdose("rtdose.dcm")     # dose grid + scaling + grid geometry
dvh = pydcm.dvhcalc("rtstruct.dcm", "rtdose.dcm", roi_number)   # ROI-for-ROI DVH
pydcm.write_rtdose(dose, affine=grid.affine, output="out.dcm")  # conformance-clean

Radiomics (IBSI)

from pydcm.radiomics import radiomics

feats = radiomics("ct.dcm", mask="roi.dcm")    # pixels → HU, spacing from geometry
feats["firstorder_Mean"], feats["glcm_Contrast"], feats["shape_VoxelVolume"]
# 135 IBSI features / 10 classes; tune bins= / resample= / distances=,
# and @pydcm.radiomics.feature("firstorder") registers your own over the same grid.

Segmentation (SEG)

pydcm.write_seg("ct.dcm", labelmap, segments, output="seg.dcm")  # coded SEG from a labelmap
labels = pydcm.read_seg("seg.dcm")               # → labelmap (voxel = Segment Number)
masks  = pydcm.read_seg("seg.dcm", masks=True)   # → per-segment binary / fractional masks

Waveforms (ECG / EEG)

from pydcm import waveforms

wf  = waveforms.read_waveform("ecg.dcm")         # leads, units, sample rate, annotations
arr = waveforms.multiplex_array("ecg.dcm")       # [channels, samples] ndarray
raw = waveforms.to_mne("eeg.dcm")                # hand straight to MNE / neurokit2

FHIR

study = pydcm.fhir.imaging_study("study_dir/")   # whole study → FHIR R4 ImagingStudy (dict)

Agent / MCP server

$ python -m pydcm.mcp        # in-process MCP server (tools/list + tools/call over stdio)

pydcm.mcp exposes 79 tools over the live engine — more than the command-line surface, because it drives in-process objects (transforms / Volume / WSI / SR & SEG authoring with write + validate). For an agent learning the toolkit, the Agent Skill (docs/skill.md) maps a task → the right pydcm call.

License

pydcm is licensed under Apache-2.0 (see LICENSE / NOTICE). The high-performance DICOM engine ships as a compiled binary inside the extension. Third-party components linked into the extension are listed in THIRD-PARTY-LICENSES — all permissive (BSD / MIT / Zlib / Apache / IJG) except FFmpeg, which is included under LGPL-2.1 (full text in LGPL-2.1.txt) for embedded-video DICOM decode, with a §6 relink offer.

pydcm distributes as wheels only (no sdist) — the engine ships as a compiled binary inside the extension, and parts of it are not open source.