Datasets:

MedOtter
/

NSCLC-Radiomics-Interobserver1

license: cc-by-nc-3.0
task_categories:
  - image-segmentation
tags:
  - medical
  - ct
  - lung
  - nsclc
  - lung-cancer
  - tumor-segmentation
  - gtv
  - inter-observer
  - radiotherapy
  - dicom
  - tcia
pretty_name: NSCLC-Radiomics-Interobserver1
size_categories:
  - n<1K
dataset_info:
  features:
    - name: patient_id
      dtype: string
    - name: ct_series_uid
      dtype: string
    - name: rt_series_uid
      dtype: string
    - name: num_ct_slices
      dtype: int32
    - name: slice_index
      dtype: int32
    - name: n_observers
      dtype: int32
    - name: consensus_voxels
      dtype: int64
    - name: mean_pairwise_dice
      dtype: float32
    - name: obs1_voxels
      dtype: int64
    - name: obs2_voxels
      dtype: int64
    - name: obs3_voxels
      dtype: int64
    - name: obs4_voxels
      dtype: int64
    - name: obs5_voxels
      dtype: int64
    - name: image
      dtype: image
    - name: mask
      dtype: image
    - name: overlay
      dtype: image
    - name: agreement
      dtype: image
  splits:
    - name: preview
      num_bytes: 5181371
      num_examples: 21
  download_size: 5201009
  dataset_size: 5181371
configs:
  - config_name: default
    data_files:
      - split: preview
        path: data/preview-*

NSCLC-Radiomics-Interobserver1

Multiple-delineation inter-observer / inter-method variability study of gross-tumour-volume (GTV) contouring on pre-treatment thoracic CT of non-small-cell lung cancer (NSCLC). For each tumour, five radiation oncologists independently delineated the GTV twice — once manually (vis) and once auto-segmentation-assisted then edited (auto) — giving up to 10 GTV delineations per patient. The collection exists specifically to quantify contouring variability, so there is no single gold-standard mask by design; all delineations are retained.

⚠️ This is NOT the main NSCLC-Radiomics ("Lung1", n=422) collection. It is the separate Interobserver1 sub-collection (22 patients) from the same Maastricht/Dana-Farber radiomics programme. It is also distinct from the RIDER-LungCT-Seg test/retest arm. See "Relationship to other collections".

Dataset Details

Field	Value
Modality	CT (pre-treatment, radiotherapy-planning thorax; mostly contrast-enhanced)
Body part	Thorax / lung
Task	3D tumour (GTV) segmentation; inter-observer variability study
Patients	22 (21 with delineations; interobs09 is CT-only)
Series	64 total — 22 CT, 21 RTSTRUCT, 21 DICOM SEG
CT slices	3,844
Observers	5 radiation oncologists (obs 1 & 3 = trainees; 2, 4, 5 = experienced)
Methods	2 per observer: `vis` (manual) and `auto` (auto-assisted + manual edit)
Format	DICOM (CT + RTSTRUCT). DICOM SEG omitted from this mirror — see below
License	CC BY-NC 3.0 Unported (Data Citation Required)
Source	The Cancer Imaging Archive (TCIA), official author upload

This HuggingFace mirror is a LEAN raw-DICOM copy: it contains the CT images (images/) and the RTSTRUCT contour objects (segmentations/). The collection's DICOM SEG objects — a rasterised duplicate of the same RTSTRUCT contours — are not included here; RTSTRUCT carries every delineation losslessly. A v3 (2020-08-31) revision of the original collection fixed an inadvertent label mismatch between the DICOM SEG and RTSTRUCT objects; this mirror was downloaded after that fix (REST API serves the current version).

Annotation structure (RTSTRUCT ROI names)

Each patient's RTSTRUCT encodes the delineations in its ROI names:

ROI name pattern	Meaning
`GTV-1vis-{1..5}`	Primary/index tumour, manual delineation by observer 1–5 — present for all 21 annotated patients
`GTV-1auto-{1..5}`	Primary tumour, auto-assisted delineation by observer 1–5 (20/21; `interobs19` has none)
`GTV-2{vis	auto}-{1..5}`
`suv2,5` / `suv_2.5`	Auxiliary PET SUV-2.5 threshold auto-contour (not an observer delineation)
`treshold0,34` / `treshhold0,34` / `tresh_34%`	Auxiliary PET 34%-SUVmax threshold auto-contour
`treshold-pr` / `treshold-ln`	Auxiliary PET threshold contour (primary / lymph node)

The auxiliary PET-threshold ROIs are part of the original radiotherapy-planning structure sets but are not the manual observer delineations and should be excluded from inter-observer analyses.

Recommended ground truth

Because the study is about variability, all observer delineations are kept. For benchmarking that needs a single reference mask, the recommended default is the STAPLE consensus of the five manual delineations of the index tumour (GTV-1vis-1 … GTV-1vis-5) — a principled probabilistic consensus across all five experts, using the pure-manual (not auto-assisted) contours, available for every annotated patient. Individual per-observer (vis/auto) contours remain available in the RTSTRUCT for variability studies; second-tumour (GTV-2*) and PET-threshold ROIs are present but excluded from the default reference.

Relationship to other collections

NSCLC-Radiomics ("Lung1", n=422) — different cohort. Interobserver1 PatientIDs use the interobsNN namespace (e.g. interobs01), disjoint from Lung1's LUNG1-xxx, and use a different CT protocol (contrast-enhanced RT-planning vs. Lung1 non-contrast). No ID-level collision. Still, dedup by PatientID / SeriesInstanceUID before any joint benchmark.
RIDER-LungCT-Seg — the test/retest arm of the same parent radiomics programme; potential shared provenance if both are used together.
series_to_patient.json preserves PatientID, SeriesInstanceUID, StudyInstanceUID, Modality, and per-series metadata for cross-referencing.

Structure

images/<PatientID>/<SeriesInstanceUID>/*.dcm          # 22 CT series
segmentations/<PatientID>/<SeriesInstanceUID>/*.dcm    # 21 RTSTRUCT (Modality=RTSTRUCT)
series_to_patient.json                                 # per-series metadata + cross-ref IDs

PatientID ranges over interobs01 … interobs33 (non-contiguous). Each RTSTRUCT references its source CT series via ReferencedFrameOfReferenceSequence → RTReferencedStudySequence → RTReferencedSeriesSequence → SeriesInstanceUID.

Splits

The collection does not prescribe train/val/test splits.

Source

TCIA collection: https://www.cancerimagingarchive.net/collection/nsclc-radiomics-interobserver1/
DOI: 10.7937/tcia.2019.cwvlpd26
Fully public — no registration required.

Citation

@misc{wee2019nsclcinterobserver1,
  author    = {Wee, Leonard and Aerts, Hugo J. W. L. and Kalendralis, Petros and Dekker, Andre},
  title     = {Data From NSCLC-Radiomics-Interobserver1 [Data set]},
  year      = {2019},
  publisher = {The Cancer Imaging Archive},
  doi       = {10.7937/tcia.2019.cwvlpd26}
}

@article{kalendralis2020fair,
  author  = {Kalendralis, Petros and Shi, Zhenwei and Traverso, Alberto and others},
  title   = {FAIR-compliant clinical, radiomics and DICOM metadata of RIDER,
             interobserver, Lung1 and head-Neck1 TCIA collections},
  journal = {Medical Physics},
  volume  = {47},
  number  = {11},
  pages   = {5931--5940},
  year    = {2020},
  doi     = {10.1002/mp.14322}
}

@article{aerts2014decoding,
  author  = {Aerts, Hugo J. W. L. and Velazquez, Emmanuel Rios and Leijenaar, Ralph T. H. and others},
  title   = {Decoding tumour phenotype by noninvasive imaging using a
             quantitative radiomics approach},
  journal = {Nature Communications},
  volume  = {5},
  pages   = {4006},
  year    = {2014},
  doi     = {10.1038/ncomms5006}
}

@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}
}