DICOM Series Preflight
Header-only validation for one DICOM series before conversion or inference.
The evidence shows least-scope directory scanning, no declared network or local/home writes, no GPU requirement, and explicit header-only, non-clinical, non-de-identification limits; Apache-2.0, NVIDIA authorship, and a security-reporting path provide attribution. Points are deducted because Bash is allowed without explicit confirmation or stronger execution isolation, outputs may expose metadata such as StudyDescription, dependency ranges are broad, and no concrete supply-chain mitigation or rollback path is documented.
The evidence shows a defined entrypoint, arguments, troubleshooting table, manifest gates, repeatability configuration, and three unit tests; abnormal inputs can produce findings and error messages. Points are deducted because nothing was executed here, key paths cannot be reproduced statically, command and manifest placeholder conventions are mildly inconsistent, and handling for permissions, very large directories, partial readability, and missing spacing is incomplete.
The evidence clearly targets developers checking one DICOM series before conversion or inference, with documented input, JSON output, scope, and non-fit scenarios, plus positive and negative evaluation tasks. Points are deducted because trigger exclusions remain general, the single-directory and CT-LPS assumptions narrow applicability, and Chinese-language interaction, host variation, and mainland-China reachability are not evidenced; the core function does not depend on overseas services.
The evidence includes SKILL.md, a skill card, manifest, output schema, script, fixtures, tests, limitations, troubleshooting, version 0.1.0, Apache-2.0 licensing, and NVIDIA ownership. Points are deducted for the missing recommended Examples section, absent explicit changelog and maintenance/update procedure, and non-pinned dependencies; repository README governance cannot fully substitute for skill-local maintenance documentation.
The evidence shows directly emitted structured JSON covering inventory, orientation, PHI flags, consistency, findings, and pass/warn/fail verdicts, with three defined evaluation tasks and unit tests. Points are deducted because this review did not execute the skill, reported benchmark numbers are not independently verified here, the scope excludes pixel-level PHI and clinical validation, and comparative marginal benefit over manual or alternative checks is only weakly evidenced.
The evidence includes revision-scoped source, schema, fixture generator, tests, manifest repeatability settings, and an evaluation report, enabling limited auditability. Points are deducted because this review did not execute anything, no skill-specific real CI evidence is shown, and the benchmark lacks independently checkable run artifacts and broad coverage; therefore the static ceiling applies.
- This is not a de-identification tool: empty standard-tag results do not exclude private tags or burnt-in pixel PHI; use a dedicated de-identifier before sharing.
- Bash execution and access to sensitive DICOM directories should be explicitly authorized by the caller and isolated; pass or warn must not be interpreted as clinical fitness or downstream segmentation assurance.
- Before deployment, review and preferably lock pydicom, nibabel, numpy, and typer versions, then run the tests and negative fixtures in the target environment.
What does this skill do, and when should you use it?
This skill performs a header-only preflight on one DICOM series directory before conversion or inference. It scans a single directory and emits inventory data, orientation axcodes, PHI flags, findings, and a pass, warn, or fail verdict. It does not decode pixels and is not intended for de-identification or clinical clearance. The documented Python entrypoint, manifest requirements, and verifier guidance define the supported workflow.
Reads a caller-provided DICOM directory and runs scripts/preflight_series.py for a single-series scan; does not decode pixel data; emits JSON containing inventory, orientation axcodes, PHI flags, findings, and preflight.verdict; and can be paired with dicom_preflight_quality_v1 for a trusted preflight pack.
- A medical-imaging engineer checks a DICOM series before converting it to another format.
- A medical-AI engineer obtains orientation, inventory, and gate results before starting inference.
- A data-pipeline developer generates structured preflight JSON for each series directory.
- A validation engineer preserves failed evidence and uses gate messages to repair inputs or wrapper code.
What are this skill's strengths and limitations?
- Produces structured JSON with pass, warn, or fail preflight verdicts.
- Reports inventory, orientation axcodes, PHI flags, and findings.
- Supports trusted evidence packs through the documented verifier and workflow.
- Requires the documented upstream entrypoint instead of a handwritten replacement.
- Header-only: it does not decode pixels or detect burnt-in PHI.
- Compressed transfer syntax and multi-frame instances are warned about but not decoded.
- Scans one directory and does not reconcile multiple studies in one tree.
- Not for de-identification, private-tag review, clinical clearance, production ingestion, or autonomous diagnosis.
- The specific Python package names in skill_manifest.yaml are not included in the supplied material.
How do you install this skill?
Install the named skill from NVIDIA's collection:
npx skills add nvidia/skills --skill dicom-series-preflight --yes
The source does not document a separate package for this skill. Reload the skills in the target agent after installation.
How do you use this skill?
From the repository root, run:
python scripts/preflight_series.py PATH_TO_DICOM_DIR
Replace PATH_TO_DICOM_DIR with one DICOM series directory and keep outputs under a caller-provided run directory. Inspect the emitted JSON and paired verifier guidance before treating the run as evidence.