extracting-dicom-metadata

Extracting DICOM Metadata & Report Text for OpenMed

DICOM (Digital Imaging and Communications in Medicine) files carry far more than pixels: a header of tagged attributes (patient, study, series, equipment) and, for DICOM-SR (Structured Reports), a content tree holding the actual radiology/cardiology report text. Two jobs sit here: pull the report narrative for NLP, and flag the PHI in the header so it gets scrubbed. This skill does both, then hands narrative to OpenMed. Header tags are read with pydicom (external, MIT-licensed); de-identification of the extracted text is OpenMed's.

When to use

• You ingest DICOM from PACS/VNA or a research archive and want the SR report text mined with clinical NLP.
• You must enumerate PHI-bearing header tags before sharing/exporting images.
• You have DICOM-SR objects (e.g. radiology measurements + impression) whose content tree contains the dictated report.

DICOM headers in one minute

Every attribute has a tag (gggg,eeee) (group, element), a VR (value representation, e.g. PN person name, DA date, UI UID), and a value. PHI clusters in well-known tags:

Quick start

Read the header, pull SR report text, flag PHI tags, hand off to OpenMed:

import pydicom
import openmed

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

# 1) Enumerate PHI-bearing header tags (report, do not log values).
PHI_TAGS = [
    (0x0010, 0x0010), (0x0010, 0x0020), (0x0010, 0x0030), (0x0010, 0x1040),
    (0x0008, 0x0090), (0x0008, 0x0050), (0x0008, 0x0020), (0x0008, 0x0030),
]
present_phi = [hex_pair for hex_pair in PHI_TAGS if hex_pair in ds]

# 2) Extract report text from a DICOM-SR content tree (recursively).
def sr_text(dataset):
    chunks = []
    for item in dataset.get("ContentSequence", []):
        vt = item.get("ValueType")
        if vt == "TEXT" and "TextValue" in item:
            chunks.append(item.TextValue)
        if "ContentSequence" in item:          # nested CONTAINER
            chunks.append(sr_text(item))
    return "\n".join(c for c in chunks if c)

report = sr_text(ds)
# Some modalities stash narrative in free-text header tags too:
for tag in ("ImageComments", "SeriesDescription", "StudyDescription"):
    if tag in ds and isinstance(ds.get(tag), str):
        report += "\n" + ds.get(tag)

# 3) De-identify the narrative, then run NER.
if report.strip():
    deid = openmed.deidentify(report, method="replace", policy="hipaa_safe_harbor")
    result = openmed.analyze_text(deid.text, output_format="dict")