TCIA 放射影像研究队列检索

SKILL.md

中文导读

检索 TCIA 公开癌症影像集合、病例、检查和 DICOM 序列元信息，估计研究数据范围与取得容量。

输入材料

• 癌种、模态与研究条件
• TCIA 公开接口访问条件

交付结果

• 影像队列元信息
• 检索条件与容量估计

阅读与使用

本页保存指定版本的原始 Skill 和配套文件，可查看中文导读、原文与参考资料。获取完整包后，按原文准备所需工具与材料。执行需要 ToolUniverse SDK 或 MCP 服务，部分数据库需要独立访问权限。医疗资料供研究、资料整理和专业人员复核，试验资格、用药与诊疗结论需要核实当前来源。

来源仓库为 mims-harvard/ToolUniverse，采用 Apache-2.0 许可。Stars 表示仓库关注度，核验日期为 2026-10-15。分类表示适用的工作环节，实际执行条件以原文为准。

原始 Skill

Discover and characterize public cancer imaging cohorts from The Cancer Imaging Archive (TCIA, National Cancer Institute) — collection, patient, study, and DICOM series-level metadata for building or scoping a radiology research cohort.

Scope: what this skill can and cannot do

Can: enumerate collections, patients, studies, and series; filter by modality/body part/manufacturer; pull DICOM series-level metadata (scanner, acquisition, license); estimate series download size; list individual DICOM instance (slice) UIDs.

Cannot: download or view actual pixel data, run segmentation/detection models, or interpret images. Every TCIA tool here returns metadata about what imaging exists — not the images themselves. If the user wants the actual pixels, tell them the DICOM download UI/API is a separate TCIA service (https://www.cancerimagingarchive.net/) outside this tool set.

TCIA data is public, de-identified cancer imaging. Study/series dates are frequently shifted to a placeholder (e.g. 2000-01-01) as part of de-identification — do not interpret these as real acquisition dates.

When to Use This Skill

• Find a public radiology imaging collection for a given cancer type or organ (e.g. lung nodule CT, glioblastoma MRI, breast MRI)
• Count patients, studies, or series in a collection, optionally filtered by modality or body part
• Check which modalities (CT/MR/PT/CR/DX/US/NM), body parts, or scanner manufacturers are represented in a collection before committing to it
• Get DICOM series-level metadata (equipment, acquisition parameters, license) for a specific SeriesInstanceUID
• Estimate total download size (bytes, image count) for a series before retrieving it
• Distinguish real patients from phantom/QA test objects in a collection

NOT for (route elsewhere):

• Microscopy/cell imaging quantification (ImageJ/CellProfiler/QuPath measurements, colony counts, fluorescence intensity) -> tooluniverse-image-analysis
• Actual pixel-level image analysis, segmentation, or diagnostic interpretation of DICOM images -> not covered by any ToolUniverse skill; these tools are metadata-only
• Tumor genomic/transcriptomic/mutation data (even for the same TCGA collections that TCIA also has imaging for) -> tooluniverse-cancer-genomics-tcga

Tool Reference

Tool	Purpose	Key params
TCIA_list_collections	List all public collection names	none
TCIA_get_patients	List patients/subjects in a collection (incl. phantom flag, species)	Collection
TCIA_get_patient_studies	List a patient's (or collection's) studies	Collection, PatientID, StudyInstanceUID
TCIA_get_series	List DICOM series (the main filterable search)	Collection, PatientID, StudyInstanceUID, Modality, BodyPartExamined
TCIA_get_modality_values	Facet: distinct modalities in a collection/body part	Collection, BodyPartExamined
TCIA_get_body_part_values	Facet: distinct body parts in a collection/modality	Collection, Modality
TCIA_get_manufacturer_values	Facet: distinct scanner manufacturers	Collection, Modality, BodyPartExamined
TCIA_get_series_metadata	Full DICOM metadata for one series	SeriesInstanceUID (required)
TCIA_get_series_size	Download size + image count for one series	SeriesInstanceUID (required)
TCIA_get_sop_instance_uids	List every individual image (slice) UID in a series	SeriesInstanceUID (required)

All 10 tools verified live against the real TCIA API. Full example calls and real (trimmed) responses: references/tcia_tool_reference.md.

Workflow

1. Find or confirm the collection. If the user names a cancer type/organ rather than a TCIA collection name, call TCIA_list_collections and match against the name pattern (TCGA collections follow TCGA-<CANCERTYPE>, e.g. TCGA-GBM, TCGA-LUAD, TCGA-BRCA; other collections have their own names, e.g. LIDC-IDRI for lung nodule CT). Confirm the exact collection name with the user before proceeding — do not guess a collection name that wasn't returned by the tool.
2. Facet the collection (optional, to characterize it before committing): TCIA_get_modality_values, TCIA_get_body_part_values, TCIA_get_manufacturer_values, each filterable by Collection.
3. Enumerate subjects. TCIA_get_patients with Collection set — check the Phantom field to exclude QA phantom objects if the user wants only real human subjects, and SpeciesDescription if a non-human collection is possible.
4. Drill into a subject's studies and series. TCIA_get_patient_studies (by Collection + PatientID) then TCIA_get_series (by Collection + PatientID, optionally narrowed by Modality/BodyPartExamined) to get SeriesInstanceUIDs.
5. Inspect or size a specific series. TCIA_get_series_metadata for full DICOM tags, TCIA_get_series_size to report download size/image count before recommending a download, TCIA_get_sop_instance_uids only if the user needs individual-slice-level references.
6. Report counts and facets actually returned by the tools — never estimate a patient/series count without calling the relevant tool.

Notes

• Collection, PatientID, Modality, and BodyPartExamined are case-sensitive and must match TCIA's exact values — get them from TCIA_list_collections / TCIA_get_modality_values / TCIA_get_body_part_values rather than guessing capitalization.
• A single patient can have multiple studies, and a single study can have many series (observed: one LIDC-IDRI patient had a 1-series study and a 9-series study) — always confirm which level (patient/study/series) the user actually wants counted.
• License varies by series (TCIA collections are typically CC-BY or CC-BY-NC) — surface the LicenseName/LicenseURI field from TCIA_get_series or TCIA_get_series_metadata if the user needs to know redistribution terms.

References

• references/tcia_tool_reference.md — all 10 tools with real example calls and real (trimmed) responses captured live against the TCIA API