QIN-LungCT-Seg

QIN multi-site collection of Lung CT data with Nodule Segmentations — a TCIA analysis result from the NCI Quantitative Imaging Network (QIN). Thoracic CT scans of non-small-cell-lung-cancer (NSCLC) patients with multi-algorithm nodule/tumor segmentations contributed by three institutions, each run three times, as a study of inter-algorithm and test-retest segmentation variability (Kalpathy-Cramer et al., J Digit Imaging 2016).

Read before benchmarking — three things make this set unusual:

1. No manual gold standard. Every mask is algorithm-generated. Each tumor is segmented 9 times (3 algorithms x 3 runs). You must define a reference yourself (e.g. STAPLE / majority consensus) — see Ground Truth below.
2. Public release is "minus-Stanford". TCIA cannot redistribute the Stanford-sourced images, so the paper's full 41 scans / 52 tumors / 468 SEG ship publicly as 31 scans / 42 tumors / 378 SEG. This mirror is the complete public set — nothing further is withheld by us.
3. Source-image leakage hazard. The CT images are drawn from other TCIA collections (LIDC-IDRI, RIDER Lung CT, QIN LUNG CT) and share their SeriesInstanceUIDs. Dedup before combining with those sets — see Overlap.

Dataset Details

Composition (by source archive)

The 409 series originate from four TCIA collections. The phantom is a physical test object (Columbia/FDA), not a patient.

There is no train/val/test split — this is a comparison/reproducibility cohort.

Ground Truth — multi-algorithm, no manual reference

All 378 masks are algorithm-generated. For every tumor there are 9 masks = 3 algorithms (Columbia / Stanford / Moffitt-USF) x 3 runs (different initial conditions; a test-retest reproducibility design). No single mask is designated the gold standard by the authors.

Each SEG's SeriesDescription encodes its provenance, e.g. QIN CT challenge[lesionNN ]algNN runM segmentation result. series_to_patient.json exposes parsed Algorithm (alg01/alg02/alg03) and Run (run1/run2/run3) fields for every SEG (the source's zero-padded run01 variants are normalised to run1). All 9 masks per tumor are preserved — choose your reference downstream:

• STAPLE consensus across the 9 (principled multi-segmentation fusion), or
• majority vote across the 9, or
• a single deterministic run (e.g. alg01/run1) as a proxy.

Tumor grouping. Single-tumor patients (LIDC-IDRI / RIDER / QIN LUNG CT) have one nodule, so PatientID identifies the tumor. The Lung Phantom holds 12 lesions in one scan; its 108 SEGs are grouped by the lesionNN token in SeriesDescription.

Cross-dataset Overlap (leakage hazard)

The 30 non-phantom CT scans are literally LIDC-IDRI / RIDER Lung CT / QIN LUNG CT series — the SeriesInstanceUID (the CT series-folder name and the SEG's ReferencedSeriesSequence) is shared with those collections. Before evaluating alongside any of:

• LIDC-IDRI and LIDC-derived sets (e.g. LUNA16),
• RIDER Lung CT and the sibling analysis result RIDER-LungCT-Seg,
• QIN LUNG CT (the separate 47-patient primary collection),

dedup by SeriesInstanceUID (reliable join key) and/or PatientID. The source Collection is recorded per series in series_to_patient.json.

Structure

images/<PatientID>/<SeriesInstanceUID>/*.dcm          # CT (31 series, 6,814 slices)
segmentations/<PatientID>/<SeriesInstanceUID>/*.dcm    # DICOM SEG (378 multiframe objects)
series_to_patient.json                                 # series-level metadata (all 409)

series_to_patient.json keys each SeriesInstanceUID to: PatientID, Collection (source archive), StudyInstanceUID, Modality, SeriesDescription, Algorithm, Run, ImageCount, FileSize, License, DOI, ThirdPartyAnalysis, and the relative path.

Important Notes for Loaders

• DICOM SEG -> labelmap conversion is needed; use pydicom-seg / dcmqi's segimage2itkimage, or parse pixel_array + PerFrameFunctionalGroupsSequence directly. Each SEG frame references its source CT slice via DerivationImageSequence -> SourceImageSequence -> ReferencedSOPInstanceUID, enabling loss-less alignment to the CT grid.
• Phantom: filter Collection == "Lung Phantom" if you want patient-only data (drops 1 scan / 12 tumors / 108 SEG).
• Mixed underlying licenses: the QIN-LUNG-CT-sourced series are CC BY 4.0, the rest CC BY 3.0 — both permissive (attribution). The analysis-result DOI is CC BY 3.0.

Source & Citation

• TCIA analysis result: https://www.cancerimagingarchive.net/analysis-result/qin-lungct-seg/
• DOI: 10.7937/k9/tcia.2015.1buvfjr7
• Official, author-deposited (Kalpathy-Cramer, Napel, Goldgof, Zhao); fully public, no registration.

@article{kalpathycramer2016lungnodule,
  author  = {Kalpathy-Cramer, Jayashree and Zhao, Binsheng and Goldgof, Dmitry and
             Gu, Yuhua and Wang, Xingwei and Yang, Hao and Tan, Yongqiang and
             Gillies, Robert and Napel, Sandy},
  title   = {A Comparison of Lung Nodule Segmentation Algorithms: Methods and
             Results from a Multi-institutional Study},
  journal = {Journal of Digital Imaging},
  volume  = {29},
  number  = {4},
  pages   = {476--487},
  year    = {2016},
  doi     = {10.1007/s10278-016-9859-z}
}

@misc{kalpathycramer2015qinlungctseg,
  author    = {Kalpathy-Cramer, J. and Napel, S. and Goldgof, D. and Zhao, B.},
  title     = {Multi-site Collection of Lung CT Data with Nodule Segmentations
               [Data set]},
  year      = {2015},
  publisher = {The Cancer Imaging Archive},
  doi       = {10.7937/k9/tcia.2015.1buvfjr7}
}

@article{clark2013tcia,
  author  = {Clark, Kenneth and Vendt, Bruce and Smith, Kirk and others},
  title   = {The Cancer Imaging Archive (TCIA): Maintaining and Operating a
             Public Information Repository},
  journal = {Journal of Digital Imaging},
  volume  = {26},
  number  = {6},
  pages   = {1045--1057},
  year    = {2013},
  doi     = {10.1007/s10278-013-9622-7}
}