DICOM Metadata Extractor
Extract selected metadata from one DICOM file and flag PHI in standard tags.
The files show a single-DICOM local read with optional output to a caller-provided JSON path, no network endpoints, no home-directory writes, and no GPU or Docker requirement. The skill clearly excludes anonymization and clinical use, making the intended data flow relatively transparent. Points are deducted because Bash is allowed, execution confirmation is not required, outputs may contain PHI without de-identification or access-control guidance, overwrite rollback is unspecified, and dependency integrity or security auditing is not evidenced.
The script, manifest, arguments, output fields, and error handling are mostly consistent; missing or unreadable inputs and DICOM read failures produce diagnosable errors, while runtime and repeatability constraints are declared. Points are deducted because static review cannot reproduce the key path, no skill-specific committed test suite is shown, and coverage of malformed DICOMs, dependency drift, output-write failures, and multiframe inputs is limited. The score therefore remains below the static ceiling of 10.
The audience, single-file input, JSON output, standard-tag PHI flag, and non-fit scenarios are clearly stated, with reasonably precise semantic triggering. The core function does not depend on an overseas online service, so no mainland-China reachability deduction applies. Points are deducted because the capability is limited to development-time metadata inspection, Chinese-language interaction is not documented, and boundaries for complex DICOMs, multiframe data, private tags, and pixel PHI have limited supporting evidence.
SKILL.md is well organized with purpose, prerequisites, script usage, limitations, troubleshooting, and examples. The manifest supplies version, inputs, outputs, dependencies, side effects, validation constraints, and cost information; the skill card identifies Apache-2.0 and NVIDIA ownership. Points are deducted for the missing recommended Examples section, unclear maintenance responsibility and changelog/update path, and reduced governance certainty because the supplied license metadata is NOASSERTION despite Apache declarations in the files.
The script directly emits structured JSON containing transfer syntax, modality, study, series, image, phi_present, and phi_tags_found. The fixture, baseline assertions, and evaluation tasks concretely cover CT modality, Study UID, PHI flags, and scope warnings, giving the core task directly usable value. Points are deducted because only a limited standard-tag subset is checked, the skill does not anonymize or detect private-tag or burnt-in pixel PHI, its output cannot by itself justify public sharing or clinical decisions, and static review cannot confirm actual runtime correctness.
The source, manifest, synthetic fixture, baseline assertions, evals, and benchmark report provide auditable material with fixed inputs, field assertions, and repeatability configuration. Points are deducted because this review has no independently reproduced execution result and no committed test suite covering the skill's key paths; the benchmark is an embedded report rather than independent verification. Evidence strength is therefore limited.
- phi_present=false does not establish that a file is safe or de-identified; private tags and burnt-in pixel PHI are not checked.
- Confirm input and output permissions before execution and avoid writing PHI-containing JSON to shared locations.
- Do not use this skill for clinical work, regulatory de-identification, diagnosis, or public-sharing safety decisions.
What does this skill do, and when should you use it?
This Agent Skill from NVIDIA’s Medical AI Skills collection reads a single DICOM file. It runs the supplied pydicom-based wrapper and reports transfer syntax, modality, grouped study, series, and image metadata, along with standard-tag PHI indicators. It is intended for metadata inspection and evidence preparation, not anonymization, clinical deployment, or clinical interpretation. Runtime arguments, dependencies, side effects, and validation gates are governed by skill_manifest.yaml.
Reads one DICOM file; runs scripts/extract_metadata.py; writes results to stdout or to a JSON file with --output; emits transfer_syntax, modality, grouped study/series/image metadata, phi_present, and phi_tags_found; and runs medagent.verifiers.dicom_metadata_quality_v1 when evidence packs undergo second-pass review.
- A medical-imaging engineer needs a quick metadata check for one DICOM file.
- A data-governance team needs to flag possible PHI in standard DICOM tags during a processing workflow.
- An evaluation or validation team needs a reproducible single-file run with reviewable evidence.
What are this skill's strengths and limitations?
- Simple interface for one DICOM input file.
- Reports grouped metadata and explicit standard-tag PHI indicator fields.
- Provides a fixed script entrypoint, optional JSON output, and an evidence-verification workflow.
- Covers only a small PS3.15-inspired standard-tag subset and is not a complete Basic Application Confidentiality Profile implementation.
- Does not check private tags, detect burnt-in pixel PHI, or provide substantial multi-frame handling.
- Explicitly unsuitable for clinical deployment, regulatory de-identification, autonomous diagnosis, or patient-facing use.
- The complete runtime dependency and argument constraints are in skill_manifest.yaml, which was not supplied here.
How do you install this skill?
Install the skill through NVIDIA/skills with: npx skills add nvidia/skills --skill dicom-metadata-extract --yes. Add --agent codex, --agent claude-code, --agent cursor, or --agent kiro-cli to target a specific client. The source does not document a separate manual folder-install procedure; install runtime dependencies according to skill_manifest.yaml.
How do you use this skill?
From the repository root, run: python scripts/extract_metadata.py PATH_TO_DICOM, or python scripts/extract_metadata.py PATH_TO_DICOM --output result.json. Keep outputs under a caller-provided run directory. For second-pass evidence review, the source provides: python -m eval_engine.run_trusted skills/dicom-metadata-extract --fixture skills/dicom-metadata-extract/fixtures/sample_ct.dcm --out runs/dicom_metadata_trusted. Do not use this skill for anonymization, private-tag review, burnt-in pixel PHI detection, or clinical interpretation.