Dicom Processing

DICOM Processing

Overview

Pipeline skill for converting, reading, and de-identifying medical images in DICOM and NIfTI formats. Encodes deterministic commands and code for:

• DICOM → NIfTI conversion with dcm2niix (including BIDS sidecars)
• Reading DICOM headers and pixel data with pydicom
• PHI removal (de-identification) with pydicom or CTP DicomAnonymizer
• NIfTI inspection with nibabel
• Batch processing grouped by SeriesInstanceUID
• BIDS-compliant directory organization

Use this skill whenever the user works with .dcm, .nii, .nii.gz, DICOM directories, or anonymization workflows.

Usage

1. Convert a DICOM series to NIfTI (dcm2niix)

# Standard conversion: gzip output, BIDS sidecar, pattern-named files
dcm2niix -z y -b y -f %p_%s -o output_dir input_dir

Flags (memorize):

Common naming tokens: %p protocol, %s series, %t time, %i patient ID, %n patient name, %d description.

2. Read a DICOM file with pydicom

import pydicom

ds = pydicom.dcmread('file.dcm')
print(ds.PatientName, ds.StudyDate, ds.Modality)
print(ds.pixel_array.shape)  # numpy array of pixel data

3. De-identify a DICOM file (pydicom)

import pydicom
from pydicom.uid import generate_uid

ds = pydicom.dcmread('file.dcm')

tags_to_remove = [
    'PatientName', 'PatientID', 'PatientBirthDate',
    'InstitutionName', 'ReferringPhysicianName', 'StudyDate',
]
for tag in tags_to_remove:
    if tag in ds:
        del ds[tag]

ds.PatientID = 'ANON_001'

# Remove private (vendor) tags — they frequently contain PHI
ds.remove_private_tags()

# Replace UIDs so the study/series/instance cannot be linked back
ds.StudyInstanceUID  = generate_uid()
ds.SeriesInstanceUID = generate_uid()
ds.SOPInstanceUID    = generate_uid()

ds.save_as('anon.dcm')

4. Batch de-identification with CTP DicomAnonymizer

# RSNA CTP / DicomBrowser DAP (DICOM Anonymizer Program)
java -jar DAP.jar -p profile.script -i input/ -o output/

Use this for large cohorts or when a validated DICOM Supplement 142 profile is required.

5. Load a NIfTI volume with nibabel

import nibabel as nib

img = nib.load('file.nii.gz')
data   = img.get_fdata()  # numpy ndarray
affine = img.affine       # 4x4 voxel-to-world transform
header = img.header

6. Batch pattern: group by SeriesInstanceUID, convert per series

from pathlib import Path
from collections import defaultdict
import pydicom, subprocess

def group_by_series(input_dir):
    series = defaultdict(list)
    for p in Path(input_dir).rglob('*.dcm'):
        ds = pydicom.dcmread(p, stop_before_pixels=True)
        series[ds.SeriesInstanceUID].append(p)
    return series

def convert_all(input_dir, output_dir):
    Path(output_dir).mkdir(parents=True, exist_ok=True)
    # dcm2niix handles series grouping internally when given a directory
    subprocess.run(
        ['dcm2niix', '-z', 'y', '-b', 'y', '-f', '%p_%s',
         '-o', output_dir, input_dir],
        check=True,
    )

7. BIDS organization

dataset/
├── dataset_description.json
├── participants.tsv
└── sub-01/
    └── ses-01/
        ├── anat/  sub-01_ses-01_T1w.nii.gz
        ├── func/  sub-01_ses-01_task-rest_bold.nii.gz
        └── dwi/   sub-01_ses-01_dwi.nii.gz

Rule: sub-<label>/ses-<label>/<modality>/<sub>_<ses>_<suffix>.nii.gz, with a matching .json sidecar.

Response Format

• Lead with the command or code the user needs — explain after
• Structure as: confirm inputs → working code → key parameters explained → gotchas
• One complete working example per task; do not show every alternative
• Keep code comments minimal and functional (what, not why-it-exists)
• Target: 50-100 lines of code with brief surrounding explanation

Core Concepts

Format and De-identification Decision Tree

What is the target format?
├─ NIfTI (analysis) → dcm2niix -z y -b y
│  ├─ Need BIDS layout? → add -f sub-%i_ses-%t_%p, reorganize post-hoc
│  └─ Single series only? → point dcm2niix at series directory directly
└─ DICOM (archive/PACS) → pydicom read/write

De-identification method:
├─ Small cohort (< 100 studies) → pydicom script (remove tags + private + replace UIDs)
├─ Large cohort or regulatory → CTP DicomAnonymizer with Supplement 142 profile
└─ Burned-in pixel PHI risk?
   ├─ US / Secondary Capture / Screenshots → YES: add OCR + pixel masking step
   └─ CT / MR / PT → Usually NO: tag-level de-ID sufficient

DICOM hierarchy. Patient → Study → Series → Instance. A "series" is the atomic unit for conversion — one NIfTI volume typically corresponds to one SeriesInstanceUID.

UIDs are identifiers, not secrets, but they are linkable. Two files sharing a StudyInstanceUID came from the same exam. De-identification must replace UIDs (generate_uid()) to break linkage across releases.

Private tags carry PHI. Vendor-specific tags (odd group numbers) often embed operator notes, patient IDs, or burned-in metadata. Always call ds.remove_private_tags().

Pixel data can contain PHI too. Ultrasound frames, secondary captures, and screenshots frequently have burned-in patient name/MRN in the pixels. Tag removal does not scrub pixels — run OCR + masking separately when this risk applies.

Key DICOM tags for geometry and modality:

These five tags are sufficient to reconstruct the NIfTI affine matrix.

NIfTI vs DICOM. NIfTI stores a single 3D/4D volume with one affine; DICOM stores slice-level metadata. Converters (dcm2niix) aggregate slices of a series and compute the affine from the tags above.

Quick Reference

# Convert
dcm2niix -z y -b y -f %p_%s -o out/ in/

# Skip merging across series (safer)
dcm2niix -z y -b y -m n -f %p_%s -o out/ in/

# Batch anonymization (CTP)
java -jar DAP.jar -p profile.script -i input/ -o output/

# Read DICOM
ds = pydicom.dcmread('file.dcm')
ds.PatientName, ds.Modality, ds.pixel_array.shape

# Header-only (fast, no pixels)
ds = pydicom.dcmread('file.dcm', stop_before_pixels=True)

# Anonymize core fields
for t in ['PatientName','PatientID','PatientBirthDate',
          'InstitutionName','ReferringPhysicianName','StudyDate']:
    if t in ds: del ds[t]
ds.remove_private_tags()
ds.StudyInstanceUID  = generate_uid()
ds.SeriesInstanceUID = generate_uid()
ds.SOPInstanceUID    = generate_uid()

# Load NIfTI
img = nib.load('file.nii.gz'); data = img.get_fdata(); aff = img.affine

Common Mistakes

• Wrong: Not removing private tags during DICOM de-identification Right: Always call ds.remove_private_tags() as part of the de-identification pipeline Why: Vendor private tags (odd group numbers) routinely contain PHI that survives standard tag-level scrubbing

• Wrong: Keeping original UIDs after anonymization Right: Replace StudyInstanceUID, SeriesInstanceUID, and SOPInstanceUID with generate_uid() for every de-identified file Why: Original UIDs are linkable back to the source PACS — two files sharing a StudyInstanceUID reveal they came from the same exam

• Wrong: Using default dcm2niix merge behavior without verification Right: Use -m n to disable merging, or inspect output file counts against the expected series list Why: Default merging can combine distinct acquisitions into one NIfTI volume, corrupting the data

• Wrong: Ignoring burned-in annotations in pixel data during de-identification Right: Add an OCR detection and pixel masking step for ultrasound, secondary capture, and screenshot modalities Why: Tag-level de-identification does not scrub pixels — patient name/MRN burned into the image persists after tag removal

• Wrong: Outputting uncompressed .nii files from dcm2niix Right: Always pass -z y to dcm2niix for gzipped .nii.gz output unless downstream tools explicitly require uncompressed Why: Uncompressed NIfTI files are 3-5× larger, wasting storage and slowing transfers with no benefit for most pipelines

• Wrong: Reading full pixel data when only DICOM headers are needed Right: Use pydicom.dcmread(path, stop_before_pixels=True) for metadata-only scans Why: Loading pixel arrays for large cohorts wastes memory and time when only tags like Modality, SeriesInstanceUID, or SliceThickness are needed

• Wrong: Treating DICOMDIR as a file directory and iterating its entries Right: Iterate actual .dcm files using Path(input_dir).rglob('*.dcm') instead Why: DICOMDIR is an index file with its own structure — it may be incomplete, outdated, or absent; direct file iteration is more reliable

References

• dcm2niix: Li et al. J Neurosci Methods 2016, https://doi.org/10.1016/j.jneumeth.2016.03.001
• pydicom: https://pydicom.github.io/pydicom/stable/
• BIDS: Gorgolewski et al. Sci Data 2016, https://doi.org/10.1038/sdata.2016.44
• DICOM de-identification: DICOM PS3.15 Annex E, https://dicom.nema.org/medical/dicom/current/output/chtml/part15/chapter_E.html