Dicom Metadata Extract

skill_manifest.yaml

# SPDX-FileCopyrightText: Copyright (c) 2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

id: medagent.dicom_metadata_extract
version: 0.1.0
upstream_refs:
  - kind: pypi_package
    name: pydicom
    version_constraint: ">=2.4.5,<4,!=3.0.0,!=3.0.1"
license: Apache-2.0
intended_use:
  summary: Engineering-time DICOM metadata extraction with a PHI-tag-presence flag.
  scope: development
  not_for:
    - clinical deployment
    - regulatory de-identification
    - autonomous diagnosis
    - patient-facing use
inputs:
  - name: dicom_path
    type: file_path
    formats: [dicom]
    max_size_bytes: 104857600
outputs:
  - name: metadata_json
    type: json
    schema: validators/output_schema.json
runtime:
  language: python
  entrypoint: scripts/extract_metadata.py
  python: ">=3.10"
  args:
    - "${python}"
    - "${script}"
    - "${fixture}"
  dependencies:
    pydicom: ">=2.4.5,<4,!=3.0.0,!=3.0.1"
    typer: ">=0.9"
  test_dependencies:
    jsonschema: ">=4.0"
  side_effects:
    pip_packages: ["pydicom>=2.4.5,<4,!=3.0.0,!=3.0.1", "typer>=0.9"]
    local_writes:
      - {path: "<caller-provided --output>", approx_mb_max: 1, optional: true}
    home_writes: []
    network_endpoints: []
    requires_docker: false
    requires_gpu: none
    env_required: []

cost:
  # Per-invocation agent-overhead token cost measured by NeMo Agent Toolkit
  # (NAT) profiler — the cost an LLM-driven agent pays to call this skill
  # once. The skill itself emits zero tokens. See tools/nat_audit/README.md
  # for methodology (pinned model, agent type, tool registry size).
  token_estimate:
    common:
      model: meta/llama-3.3-70b-instruct
      agent_type: tool_calling_agent
      measured_at: 2026-05-16
      methodology: tools/nat_audit/README.md
    isolated_tool_call:
      prompt_tokens: 2250
      completion_tokens: 40
      total_tokens: 2290
      llm_calls: 2
      n_tools_in_workflow: 9
    end_to_end_workflow:
      prompt_tokens: 4810
      completion_tokens: 83
      total_tokens: 4893
      llm_calls: 3
      n_tools_in_workflow: 11
      scenario: realistic_user_workflow

paired_verifiers:
  - id: medagent.verifiers.dicom_metadata_quality_v1
    status: implemented
    consumes: evidence_pack_dir
    purpose: >
      Second-pass metadata evidence audit. Confirms the source pack passed,
      required header facts are present, PHI flag and tag-list semantics are
      consistent, and the limited PHI scope is disclosed. Standard PHI tag
      presence is warning-only because this skill does not de-identify data.

limitations:
  - Small PS3.15-inspired standard-tag subset only; not a complete Basic
    Application Confidentiality Profile implementation.
  - Private tags not checked
  - Burnt-in pixel PHI not detected
  - Multi-frame handling minimal
phi_scope_disclaimer: >
  This skill flags PHI presence in a small standard DICOM tag subset only.
  It is NOT a de-identifier and does NOT remove or modify data. Private
  tags and burnt-in pixel text are out of scope.
validation:
  expected_runtime_seconds:
    max: 5.0
  sanity_checks:
    # The extractor must always populate Modality and the PHI-presence
    # flag from a readable DICOM. matches: ".+" requires non-empty.
    - {path: modality, matches: ".+"}
    - {path: phi_present, exists: true}
    - {path: phi_scope_disclaimer, matches: "Standard DICOM"}
    - {path: study.StudyInstanceUID, matches: ".+"}
    - {path: image.Rows, gt: 0}
    - {path: image.Columns, gt: 0}
  expected_cost:
    # Pydicom header read on a single CT slice is essentially I/O bound.
    # Bounds are intentionally generous — the canonical silent-failure
    # detector here is the runtime envelope, not cost. cpu_seconds.max is
    # a soft fence for "interpreter spun up but did real work".
    wall_seconds:        {max: 5}
    cpu_seconds:         {max: 5}
    rss_mb_peak:         {max: 250}
    gpu_seconds:         {max: 0}
    gpu_memory_mb_peak:  {max: 0}
  env_pin:
    # PEP 440 constraints on upstream tools whose behaviour this skill
    # depends on. The eval_engine's env_pin gate fails the run if any
    # installed version drifts outside these bounds — catches what the
    # loose `runtime.dependencies` ranges silently allow (e.g. a pydicom
    # 3 → 4 bump that changes the save_as API).
    pydicom: ">=2.4.5,<4,!=3.0.0,!=3.0.1"
  reproducibility:
    mode: repeat
    fixture: fixtures/sample_ct.dcm
    fixture_builder: fixtures/generate_sample.py
    runs: 2