Mediastinal-Lymph-Node-SEG (LNQ 2023)

Mediastinal lymph node quantification dataset from the LNQ 2023 challenge: 513 chest CT scans with manual radiologist segmentations of mediastinal lymph nodes, distributed under two annotation regimes — partial (weakly-supervised) for training, and full (gold-standard) for evaluation.

Dataset Details

Field	Value
Modality	CT (chest)
Body part	Mediastinum (chest), lymph nodes
Task	3D binary segmentation (foreground = mediastinal lymph nodes)
Patients	513 (274 F / 239 M)
CT series	513
SEG series	513 (one per patient)
Annotation regimes	"Partially Annotated" (393) + "Fully Annotated" (120)
Images	~66,400 CT DICOM slices + 513 DICOM SEG objects
Format	DICOM (images) + DICOM SEG (segmentations)
License	CC BY 4.0

Annotation Regimes

The collection contains two distinct mask regimes, both produced in 3D Slicer 4.11 by trained radiologists with axial delineation and sagittal/coronal review. The regime is stored on the SEG SOP as SeriesDescription:

1. Partially Annotated (393 cases) — only the clinically-tracked target lymph node(s) per scan are delineated, following clinical-trial bidimensional measurement protocols. The selected lesion(s) may not be the largest in the scan. Intended use: weakly-supervised training input.
2. Fully Annotated (120 cases) — every clinically-relevant mediastinal lymph node with short-axis ≥1 cm is delineated, capturing the full disease burden. This corresponds to the LNQ 2023 challenge validation (20 cases) + held-out test (100 cases) splits combined.

Recommended ground truth: the Fully Annotated cases. The MELBA challenge paper designates these as the gold-standard reference; the Partially Annotated masks are intentionally incomplete and unsuitable as a definitive reference. Evaluation in the challenge (Dice, Average Symmetric Surface Distance) is computed against the full masks.

Structure

images/<PatientID>/<StudyInstanceUID>/<SeriesInstanceUID>/*.dcm        # CT
segmentations/<PatientID>/<StudyInstanceUID>/<SeriesInstanceUID>/*.dcm # DICOM SEG
series_to_patient.json                                                 # series-level metadata

PatientID follows the pattern case_NNNN (e.g. case_0140). The DICOM SEG carries SeriesDescription ∈ {Partially Annotated, Fully Annotated} which determines the annotation regime, and references its source CT series via ReferencedSeriesSequence (top-level CT SeriesInstanceUID) and PerFrameFunctionalGroupsSequence → DerivationImageSequence → SourceImageSequence (per-frame source CT SOPInstanceUID), enabling loss-less alignment to the CT slice grid.

series_to_patient.json contains the SeriesDescription field for every SEG, so loaders can split partial-vs-full without reading DICOM headers.

Important Notes for Loaders

• DICOM SEG ⇄ ITK conversion is needed to get a labelmap volume; use dcmqi's segimage2itkimage or pydicom-seg.
• Not all 393 partial cases segment the largest lesion. Treat partial masks as positive-pixel evidence only; absence of a label outside the delineated lymph node is not a true-negative.
• Splits: if you need the original challenge train/val/test split, use SeriesDescription. The 393 partial cases form the LNQ 2023 training set; the 120 fully-annotated cases are the LNQ 2023 val (20) + test (100) sets combined — TCIA does not redistribute the LNQ-internal val/test boundary.

Source

• TCIA collection: https://www.cancerimagingarchive.net/collection/mediastinal-lymph-node-seg/
• DOI: 10.7937/QVAZ-JA09
• LNQ 2023 challenge: https://lnq2023.grand-challenge.org/
• Released: 2024-08-28 (now fully public, no registration required since 2025-07-07)

Citation

@misc{idris2024mediastinallnq,
  author    = {Idris, Tagwa and Somarouthu, Bhanusupriya and Jacene, Heather and
               LaCasce, Ann and Ziegler, Erik and Pieper, Steve and Khajavi, Reza
               and Dorent, Reuben and Pujol, Sonia and Kikinis, Ron and
               Harris, Gordon},
  title     = {Mediastinal Lymph Node Quantification (LNQ): Segmentation of
               Heterogeneous CT Data [Dataset]},
  year      = {2024},
  publisher = {The Cancer Imaging Archive},
  doi       = {10.7937/QVAZ-JA09}
}

@article{dorent2025lnq2023,
  author  = {Dorent, Reuben and Khajavi, Reza and Idris, Tagwa and others},
  title   = {{LNQ} 2023 challenge: Benchmark of weakly-supervised techniques
             for mediastinal lymph node quantification},
  journal = {Machine Learning for Biomedical Imaging},
  volume  = {3},
  year    = {2025},
  doi     = {10.59275/j.melba.2025-d482}
}