176 chest + abdomen CT scans with manually-traced voxel-wise mediastinal and abdominal lymph node segmentations. The collection underpins the Roth 2014 detection benchmark and Seff 2015 segmentation benchmark, and remains a widely-cited reference for thoracic-abdominal lymph node CADe work.

Dataset Details

Field	Value
Modality	CT
Body part	Mediastinum + Abdomen
Task	3D binary segmentation (foreground = lymph nodes)
Patients	176 (90 mediastinal + 86 abdominal)
Studies	176
CT series	176
SEG series (V5 DICOM-SEG)	176 (one per patient)
NIfTI masks (V4)	176 (one per patient, in `supplementary/MED_ABD_LYMPH_MASKS.zip`)
Images (slices)	110,179 CT DICOM slices
Annotated nodes	983 (388 mediastinal + 595 abdominal)
Format	DICOM (CT, DICOM-SEG) + NIfTI (legacy masks)
License	CC BY 3.0

Anatomical Subsets

The collection is partitioned by anatomical region (no ML train/val/test split is prescribed). Patient ID prefix encodes the subset:

Subset	Patients	Annotated nodes	Measurement convention
*`MED_LYMPH_`**	90	388	Shortest axis only (RECIST)
*`ABD_LYMPH_`**	86	595	Longest + shortest axis (axial view)

Mask Sources

The collection ships three companion annotation archives. MASKS is the recommended ground truth; the other two are pre-existing detector inputs from the original CADe pipeline and are retained here for reproducibility.

Source	Role	Path on HF
DICOM-SEG (V5, 2023-03-31) ★	Recommended GT — per-patient binary lymph-node masks, aligned to source CT via per-frame `DerivationImageSequence`.	`segmentations/<PatientID>/<StudyInstanceUID>/<SeriesInstanceUID>/*.dcm`
NIfTI masks (V4, 2015-12-14)	Same voxel-wise GT as DICOM-SEG, expressed in NIfTI.	`supplementary/MED_ABD_LYMPH_MASKS.zip`
Annotations (centroids)	Lymph-node centroid points and RECIST size measurements (`.mps`, `.txt`).	`supplementary/MED_ABD_LYMPH_ANNOTATIONS.zip`
Candidates (CADe detections)	Computer-generated positive/negative candidate lists from Cherry/Liu SPIE 2014 — detector proposals, not ground truth.	`supplementary/MED_ABD_LYMPH_CANDIDATES.zip`

Recommended GT: DICOM-SEG masks (V5). TCIA wiki explicitly directs citing Seff 2015 (MICCAI 2015, DOI 10.1007/978-3-319-24571-3_7) when using these masks. The DICOM-SEG representation was added in V5 (2023) and aligns to each source CT slice via standard DICOM cross-references. The NIfTI masks (V4) contain the same delineations and are also provided for legacy compatibility.

The TCIA wiki notes the centroid annotations and the masks were produced independently and their indexing is not aligned — a loader that needs both must key on patient ID, not on node index within a patient.

Structure

images/<PatientID>/<StudyInstanceUID>/<SeriesInstanceUID>/*.dcm        # CT
segmentations/<PatientID>/<StudyInstanceUID>/<SeriesInstanceUID>/*.dcm # DICOM-SEG
supplementary/MED_ABD_LYMPH_MASKS.zip                                  # NIfTI masks
supplementary/MED_ABD_LYMPH_ANNOTATIONS.zip                            # centroid .mps files
supplementary/MED_ABD_LYMPH_CANDIDATES.zip                             # CADe proposals
series_to_patient.json                                                 # series-level metadata

PatientID follows the pattern MED_LYMPH_NNN (90 cases) or ABD_LYMPH_NNN (86 cases). The DICOM-SEG references its source CT via ReferencedSeriesSequence (top-level CT SeriesInstanceUID) and PerFrameFunctionalGroupsSequence -> DerivationImageSequence -> SourceImageSequence (per-frame source CT SOPInstanceUID).

series_to_patient.json lists every CT and SEG series with PatientID, StudyInstanceUID, SeriesInstanceUID, Modality, SeriesDescription, ImageCount, FileSize, and the on-disk relative path, so loaders can index patients without crawling the DICOM headers.

Notes for Loaders

DICOM-SEG -> labelmap: use pydicom-seg or dcmqi's segimage2itkimage for an aligned 3D label volume. Alternatively, load the NIfTI from supplementary/MED_ABD_LYMPH_MASKS.zip and resample it to the CT grid.
No prescribed split: a single train split covering all 176 patients. Downstream tasks may carve out evaluation patients by patient ID; the anatomical subset (MED vs ABD) is the natural stratification axis.
Subset detection: the patient-ID prefix (MED_ vs ABD_) is the authoritative subset label.

Source

TCIA collection: https://www.cancerimagingarchive.net/collection/ct-lymph-nodes/
DOI: 10.7937/K9/TCIA.2015.AQIIDCNM
Released: V5 on 2023-03-31 (DICOM-SEG version of masks added). Fully public, no registration required since 2025-07-07.

Citation

@misc{roth2015ctlymphnodes,
  author    = {Roth, Holger and Lu, Le and Seff, Ari and Cherry, Kevin M. and
               Hoffman, Joanne and Wang, Shijun and Liu, Jiamin and
               Turkbey, Evrim and Summers, Ronald M.},
  title     = {A new 2.5D representation for lymph node detection in CT [Dataset]},
  year      = {2015},
  publisher = {The Cancer Imaging Archive},
  doi       = {10.7937/K9/TCIA.2015.AQIIDCNM}
}

@inproceedings{roth2014lymphnode,
  author    = {Roth, Holger R. and Lu, Le and Seff, Ari and Cherry, Kevin M. and
               Hoffman, Joanne and Wang, Shijun and Liu, Jiamin and
               Turkbey, Evrim and Summers, Ronald M.},
  title     = {A New {2.5D} Representation for Lymph Node Detection Using
               Random Sets of Deep Convolutional Neural Network Observations},
  booktitle = {Medical Image Computing and Computer-Assisted Intervention --
               {MICCAI} 2014},
  pages     = {520--527},
  year      = {2014},
  doi       = {10.1007/978-3-319-10404-1_65}
}

@inproceedings{seff2015lymphnodemasks,
  author    = {Seff, Ari and Lu, Le and Barbu, Adrian and Roth, Holger and
               Shin, Hoo-Chang and Summers, Ronald M.},
  title     = {Leveraging Mid-Level Semantic Boundary Cues for Automated
               Lymph Node Detection},
  booktitle = {Medical Image Computing and Computer-Assisted Intervention --
               {MICCAI} 2015},
  year      = {2015},
  doi       = {10.1007/978-3-319-24571-3_7}
}

Downloads last month: 2,290

Total file size:

58.4 GB