dicom-processing（DICOM 处理）

DICOM Processing

What This Skill Does

Generates correct code for reading, writing, and manipulating DICOM (Digital Imaging and Communications in Medicine) files. Covers the pydicom Python library (primary), DCMTK command-line tools, and the DICOM data model including tags, VRs, transfer syntaxes, and SOP classes.

Prerequisites

• Python 3.8+ with pydicom (pip install pydicom)
• Optional: pillow or numpy for pixel data operations
• Optional: pylibjpeg + pylibjpeg-libjpeg for compressed transfer syntaxes
• Optional: DCMTK toolkit for command-line operations

# Core
pip install pydicom

# Pixel data handling
pip install pydicom[all]
# Or individually:
pip install numpy pillow pylibjpeg pylibjpeg-libjpeg pylibjpeg-openjpeg

# DCMTK (command-line tools)
# macOS
brew install dcmtk
# Ubuntu/Debian
apt-get install dcmtk

Quick Start

Read a DICOM File

import pydicom

ds = pydicom.dcmread("image.dcm")

# Access common attributes
print(ds.PatientName)        # Patient's name
print(ds.StudyDate)          # Study date (YYYYMMDD)
print(ds.Modality)           # CT, MR, US, CR, etc.
print(ds.StudyInstanceUID)   # Unique study identifier
print(ds.SeriesInstanceUID)  # Unique series identifier
print(ds.SOPInstanceUID)     # Unique instance identifier

Modify and Save

ds = pydicom.dcmread("image.dcm")
ds.PatientName = "ANONYMOUS"
ds.PatientID = "ANON001"
ds.save_as("modified.dcm")

Access Pixel Data

ds = pydicom.dcmread("image.dcm")
pixel_array = ds.pixel_array  # Returns numpy ndarray
print(pixel_array.shape)      # e.g., (512, 512) for a single frame
print(pixel_array.dtype)      # e.g., int16

DICOM Data Model

Tags

Every DICOM attribute is identified by a (group, element) tag pair:

from pydicom.tag import Tag

# Access by keyword (preferred)
ds.PatientName

# Access by tag number
ds[0x0010, 0x0010]  # Same as PatientName

# Access by Tag object
ds[Tag(0x0010, 0x0010)]

# Check if tag exists
if "PatientName" in ds:
    print(ds.PatientName)

Common Tags Reference

Value Representations (VRs)

VRs define the data type and format of a DICOM value:

Sequences

Sequences are nested datasets (like arrays of objects):

# Read a sequence
if "ReferencedStudySequence" in ds:
    for item in ds.ReferencedStudySequence:
        print(item.ReferencedSOPClassUID)
        print(item.ReferencedSOPInstanceUID)

# Create a sequence
from pydicom.dataset import Dataset
from pydicom.sequence import Sequence

item = Dataset()
item.ReferencedSOPClassUID = "1.2.840.10008.5.1.4.1.1.2"
item.ReferencedSOPInstanceUID = pydicom.uid.generate_uid()

ds.ReferencedStudySequence = Sequence([item])

Working with Pixel Data

Read Pixel Data as NumPy Array

import pydicom
import numpy as np

ds = pydicom.dcmread("ct_image.dcm")
pixels = ds.pixel_array  # numpy ndarray

# Apply rescale slope/intercept for CT (Hounsfield units)
if hasattr(ds, "RescaleSlope") and hasattr(ds, "RescaleIntercept"):
    hu = pixels * ds.RescaleSlope + ds.RescaleIntercept

Window/Level for Display

def apply_window(pixels, window_center, window_width):
    """Apply window/level to pixel data for display."""
    img_min = window_center - window_width // 2
    img_max = window_center + window_width // 2
    windowed = np.clip(pixels, img_min, img_max)
    windowed = ((windowed - img_min) / (img_max - img_min) * 255)
    return windowed.astype(np.uint8)

# Common CT windows
LUNG_WINDOW = (-600, 1500)      # center, width
BONE_WINDOW = (400, 1800)
SOFT_TISSUE = (40, 400)
BRAIN_WINDOW = (40, 80)

display = apply_window(hu, *SOFT_TISSUE)

Save as PNG

from PIL import Image

# For grayscale (CT, MR, CR)
img = Image.fromarray(display, mode="L")
img.save("output.png")

# For RGB (ultrasound, pathology)
if ds.PhotometricInterpretation == "RGB":
    img = Image.fromarray(pixels, mode="RGB")
    img.save("output.png")

Multi-frame Images

ds = pydicom.dcmread("multiframe.dcm")
pixels = ds.pixel_array  # Shape: (num_frames, rows, cols)

print(f"Frames: {ds.NumberOfFrames}")
print(f"Shape: {pixels.shape}")

# Access individual frames
frame_0 = pixels[0]

Transfer Syntaxes

Transfer syntaxes define how DICOM data is encoded (byte order, compression):

print(ds.file_meta.TransferSyntaxUID)

Decompressing Pixel Data

# Install handlers for compressed transfer syntaxes
# pip install pylibjpeg pylibjpeg-libjpeg pylibjpeg-openjpeg

ds = pydicom.dcmread("compressed.dcm")
ds.decompress()  # Convert to uncompressed in-memory
pixels = ds.pixel_array

Converting Transfer Syntax

# Using DCMTK
# Decompress to Explicit VR Little Endian
dcmconv +te input.dcm output.dcm

# Compress to JPEG 2000 Lossless
dcmcjp2k +e2 input.dcm output.dcm

# Compress to JPEG Lossless
dcmcjpls +el input.dcm output.dcm

Creating DICOM Files

Create a DICOM File from Scratch

import pydicom
from pydicom.dataset import Dataset, FileDataset
from pydicom.uid import generate_uid, ExplicitVRLittleEndian
from pydicom.sequence import Sequence
import numpy as np
import datetime

# Create file dataset
filename = "new_image.dcm"
file_meta = pydicom.Dataset()
file_meta.MediaStorageSOPClassUID = "1.2.840.10008.5.1.4.1.1.2"  # CT
file_meta.MediaStorageSOPInstanceUID = generate_uid()
file_meta.TransferSyntaxUID = ExplicitVRLittleEndian

ds = FileDataset(filename, {}, file_meta=file_meta, preamble=b"\x00" * 128)

# Patient info
ds.PatientName = "Test^Patient"
ds.PatientID = "TEST001"
ds.PatientBirthDate = "19900101"
ds.PatientSex = "O"

# Study info
ds.StudyInstanceUID = generate_uid()
ds.StudyDate = datetime.date.today().strftime("%Y%m%d")
ds.StudyTime = datetime.datetime.now().strftime("%H%M%S")
ds.Modality = "CT"

# Series info
ds.SeriesInstanceUID = generate_uid()
ds.SeriesNumber = 1

# Instance info
ds.SOPClassUID = "1.2.840.10008.5.1.4.1.1.2"
ds.SOPInstanceUID = file_meta.MediaStorageSOPInstanceUID
ds.InstanceNumber = 1

# Image info
ds.Rows = 512
ds.Columns = 512
ds.BitsAllocated = 16
ds.BitsStored = 16
ds.HighBit = 15
ds.PixelRepresentation = 1  # signed
ds.SamplesPerPixel = 1
ds.PhotometricInterpretation = "MONOCHROME2"

# Pixel data
pixel_data = np.zeros((512, 512), dtype=np.int16)
ds.PixelData = pixel_data.tobytes()

ds.save_as(filename)

Bulk Operations

Iterate DICOM Files in a Directory

from pathlib import Path
import pydicom

def iter_dicom_files(directory: str):
    """Yield (path, dataset) for all DICOM files in a directory tree."""
    for path in Path(directory).rglob("*"):
        if path.is_file():
            try:
                ds = pydicom.dcmread(str(path), stop_before_pixels=True)
                yield path, ds
            except pydicom.errors.InvalidDicomError:
                continue

# Extract metadata from all files
for path, ds in iter_dicom_files("/data/studies"):
    print(f"{path}: {ds.PatientName} | {ds.Modality} | {ds.StudyDate}")

Group Files by Study/Series

from collections import defaultdict

studies = defaultdict(lambda: defaultdict(list))

for path, ds in iter_dicom_files("/data/incoming"):
    study_uid = ds.StudyInstanceUID
    series_uid = ds.SeriesInstanceUID
    studies[study_uid][series_uid].append(path)

for study_uid, series in studies.items():
    print(f"Study {study_uid}: {len(series)} series")
    for series_uid, files in series.items():
        print(f"  Series {series_uid}: {len(files)} instances")

Read Metadata Only (Fast)

# stop_before_pixels=True skips pixel data -- much faster for metadata-only operations
ds = pydicom.dcmread("large_image.dcm", stop_before_pixels=True)

Extract Specific Tags

# Read only specific tags (fastest for large datasets)
ds = pydicom.dcmread("image.dcm", specific_tags=[
    "PatientName", "PatientID", "StudyDate", "Modality",
    "StudyInstanceUID", "SeriesInstanceUID",
])

DCMTK Command-Line Tools

Common Commands

# Dump DICOM metadata
dcmdump image.dcm

# Dump specific tags
dcmdump +P "0010,0010" +P "0008,0060" image.dcm

# Modify tags
dcmodify -m "(0010,0010)=ANONYMOUS" image.dcm

# Convert transfer syntax
dcmconv +te input.dcm output.dcm    # To Explicit VR LE

# Validate DICOM conformance
dcmpschk image.dcm

# Send via C-STORE
storescu -v -aec REMOTE_AE host port image.dcm

# Query via C-FIND
findscu -v -aec REMOTE_AE host port -k "0008,0060=CT" -k "0010,0010=Smith*"

# Retrieve via C-MOVE
movescu -v -aec REMOTE_AE -aem MY_AE host port -k "0020,000D=1.2.3..."

Gotchas

• PatientName uses ^ as separator: "Family^Given^Middle^Prefix^Suffix". Use str(ds.PatientName) for display, ds.PatientName.family_name for components.
• Dates are strings, not date objects: StudyDate is "20250115", not a Python date. Parse with datetime.strptime(ds.StudyDate, "%Y%m%d").
• UIDs must be globally unique: Always use pydicom.uid.generate_uid() when creating new studies/series/instances. Never reuse UIDs.
• Pixel data may be compressed: Always handle the case where ds.pixel_array raises an error due to missing decompression handlers. Install pylibjpeg packages.
• Private tags: Vendor-specific data uses odd group numbers (e.g., (0009,xxxx)). Access with ds[0x0009, 0x0010].
• Encoding: DICOM defaults to ISO-IR 100 (Latin-1). Check SpecificCharacterSet for non-Latin text. pydicom handles decoding automatically.
• File vs dataset: Use pydicom.dcmread() to read files. The returned FileDataset includes file meta information. For in-memory datasets, use Dataset() directly.
• Modifying PixelData: If you modify pixel data, update Rows, Columns, BitsAllocated, BitsStored, HighBit, PixelRepresentation, and PhotometricInterpretation to match.

Resources

• pydicom Documentation
• pydicom User Guide
• DICOM Standard Browser
• DCMTK Documentation
• DICOM Standard (official)