Datasets:

MedOtter
/

NSCLC-PleuralEffusion

Voxel-level thoracic-cavity and pleural-effusion segmentations on the NSCLC-Radiomics CT collection. Published by Kiser et al. (Medical Physics 2020) as PleThora, "Pleural effusion and thoracic cavity segmentations in diseased lungs for benchmarking chest CT processing pipelines."

Dataset Details

Field	Value
Modality	CT (chest, contrast and non-contrast mixed)
Body part	Chest — thoracic cavity, pleural effusion
Tasks	Thoracic-cavity 3D segmentation; pleural-effusion 3D segmentation
CT patients	402 (plus 3 extras for mask-coverage: 405 total CT folders)
Thoracic-cavity masks	402 patients (primary reviewer) + 86 (secondary reviewer)
Pleural-effusion masks	78 patients (first + second reviewer) + 16 (third reviewer)
DICOM slices	~48,568
Mask format	NIfTI (.nii.gz)
CT format	DICOM
License	CC BY-NC 3.0 (NSCLC-Radiomics CTs); PleThora masks released under CC BY 3.0. The most restrictive of these (CC BY-NC 3.0) governs the combined upload.

Subsets and Recommended Ground Truth

PleThora ships multiple reviewer tiers for both subsets. The paper's benchmark splits and recommended references are:

Thoracic Cavity

Whole-cavity volume — includes lung, tumor, atelectasis and adhesions. Not a lung-parenchyma mask. The upstream description also claims it covers any effusion inside the cavity; measured against the effusion masks it does not (median 75% coverage) — see "The two tasks PARTIALLY overlap" below.

Label encoding — these masks are NOT binary. Every thoracic-cavity file carries values {0, 1, 2}: 1 = left hemithorax, 2 = right hemithorax. Verified on all 402 patients; the files are stored ('L','A','S') so the higher-x component is the patient's left. Binarize with > 0 for a whole-cavity task, or keep the two labels for a left/right task. Reading these as binary silently discards the laterality split.

Reviewer	Cases	Role
`primary_reviewer`	402	Medical-student manual correction of an initial U-Net pre-segmentation. Used as training-set GT in the paper's baseline.
`secondary_reviewer`	86	Radiologist/expert revision. Recommended GT where available.
`primary_reviewer_with_nodal_conglomerate`	1	Special variant for a single case where nodal conglomerate was treated separately.

Recommended dataloader behavior: use secondary_reviewer when present (86 paper test cases); fall back to primary_reviewer for the remaining 316.

Pleural Effusion (binary effusion mask)

Reviewer	Cases	Role
`first_reviewer`	78	Medical-student manual delineation.
`second_reviewer`	78	Radiologist revision. Recommended GT.
`third_reviewer`	16	Second radiologist for inter-observer agreement.

Recommended dataloader behavior: always use second_reviewer.

Effusion masks are binary {0, 1} for 77 of the 78 patients. The one exception is LUNG1-205_effusion_second_reviewer.nii.gz, which carries {0, 1, 2}; binarize with > 0.

A patient with no effusion file has no effusion — that is a real negative, not missing ground truth. Only 78 of the 405 patients developed one. Do not drop the other 327 as unlabelled if you are training a detector.

The two tasks PARTIALLY overlap — neither nested nor disjoint

The prose description says the cavity contains any effusion within it, but that does not hold in the data. Measured over all 78 effusion patients (second_reviewer effusion vs primary_reviewer cavity, identical headers so no resampling is involved), the fraction of effusion voxels lying inside the cavity mask is:

min	p10	median	p90	max
6.5%	27.9%	75.4%	93.6%	97.4%

No patient is ≥99% contained, and 14 of 78 are under 50% (worst: LUNG1-107 6.5%, LUNG1-013 9.3%, LUNG1-051 10%). This is consistent with how the cavity masks were made — a U-Net trained on chest CTs without cancer, then hand-corrected — so effusion-filled pleural space is often outside the cavity contour.

Consequences: the two labels cannot be merged into one {0, 1, 2} map (they are not disjoint, so one would erase the other where they meet, and they are not nested either, so a union cannot recover the superset). Treat them as two independent binary tasks / channels. Do not compute one from the other.

Splits

TCIA does not publish a fixed train/val/test split. The paper's baseline uses the 316 cases with only primary_reviewer thoracic masks for training and the 86 cases with secondary_reviewer thoracic masks for testing. We do not re-shard the data; all patients are placed under images/ and consumers can follow the paper's split using the reviewer tiers above.

Structure

images/<PatientID>/*.dcm                                 # CT DICOM series, one folder per patient (405 folders)
masks/thoracic_cavity/<PatientID>/<PatientID>_<tier>.nii.gz
masks/pleural_effusion/<PatientID>/<PatientID>_<tier>.nii.gz
manifest.json                                            # Per-patient summary: CT slice count, mask tier availability
NSCLC-Radiomics-OriginalCTs.tcia                         # Original PleThora .tcia manifest (provenance)

<tier> is one of thor_cav_primary_reviewer, thor_cav_secondary_reviewer, thor_cav_primary_reviewer_with_nodal_conglomerate, effusion_first_reviewer, effusion_second_reviewer, or effusion_third_reviewer.

Known Mask/CT Mismatches

Three patients in the PleThora mask zips are NOT in the PleThora .tcia CT manifest: LUNG1-083, LUNG1-095, LUNG1-246. We located the matching CT series for each in the broader NSCLC-Radiomics collection and added them to images/ so all mask patients have a paired CT. Three patients are in the CT manifest but have no thoracic-cavity mask: LUNG1-198, LUNG1-203, LUNG1-204. The dataloader should iterate the mask folders, not the CT folders, to skip orphan CTs.

Overlap With Other Datasets — read before benchmarking

Every CT here comes from NSCLC-Radiomics (Lung1): 402 of that collection's 422 patients, plus 3 mask-only patients recovered from the same collection. Patient IDs are kept in their original LUNG1-xxx form precisely so they can be excluded downstream — do not renumber them.

Known consequences:

nsclc-radiomics (this org) is partly derived from PleThora. Its concat_label.nii.gz is 2-channel, and channel 0 is the union of this dataset's thoracic-cavity mask — Dice 1.0000 on the patients checked. Its 85 patients are therefore already a PleThora thoracic subset under another name.
Disjoint (safe to evaluate alongside): NSCLC-Radiomics-Interobserver1 is a separate cohort of the 1,019-patient Aerts series, not a Lung1 subset; MSD Task06 Lung is drawn from NSCLC Radiogenomics (Stanford), a different collection; RIDER-LungCT, QIN-LungCT, LIDC-IDRI and 4D-Lung are unrelated institutions.
LCTSC S1 is unresolved. S1 is MAASTRO — the same institution as Lung1 — but no public cross-reference exists and byte-level slice comparison found no overlap. Treat as unknown risk on ≤20 cases.

Note that TCIA re-hashes DICOM UIDs per submission, so SeriesInstanceUID / StudyInstanceUID collisions are not a valid overlap test across collections. Match on the LUNG1-xxx patient ID instead.

Sources

TCIA analysis-result page: https://www.cancerimagingarchive.net/analysis-result/plethora/
Parent CT collection: https://www.cancerimagingarchive.net/collection/nsclc-radiomics/
TCIA DOI (PleThora): https://doi.org/10.7937/tcia.2020.6c7y-gq39

Citation

@article{kiser2020plethora,
  author  = {Kiser, Kendall J. and Ahmed, Sara and Stieb, Sonja and
             Mohamed, Abdallah S. R. and Elhalawani, Hesham and
             Park, Peter Y. S. and Doyle, Nicolette S. and Wang, Brian J. and
             Barman, Arpan and Li, Zhenyu and Cheng, Wesley and
             Anderjaska, James and Fuller, Clifton D. and Frank, Steven J. and
             Lai, Stephen Y. and Marai, G. Elisabeta and Gunn, G. Brandon and
             Garden, Adam S. and Rosenthal, David I. and Jaffray, David A. and
             Court, Laurence E. and Aerts, Hugo J. W. L.},
  title   = {PleThora: Pleural effusion and thoracic cavity segmentations in
             diseased lungs for benchmarking chest CT processing pipelines},
  journal = {Medical Physics},
  volume  = {47},
  number  = {11},
  pages   = {5941--5952},
  year    = {2020},
  doi     = {10.1002/mp.14424}
}

@misc{plethora2020tcia,
  author    = {Kiser, K. J. and Ahmed, S. and Stieb, S. and others},
  title     = {PleThora: Pleural effusion and thoracic cavity segmentations
               in diseased lungs for benchmarking chest CT processing pipelines
               [Dataset]},
  year      = {2020},
  publisher = {The Cancer Imaging Archive},
  doi       = {10.7937/tcia.2020.6c7y-gq39}
}

@article{aerts2014decoding,
  author  = {Aerts, Hugo J. W. L. and Velazquez, Emmanuel Rios and
             Leijenaar, Ralph T. H. and Parmar, Chintan and Grossmann, Patrick
             and Carvalho, Sara and Bussink, Johan and Monshouwer, Rene and
             Haibe-Kains, Benjamin and Rietveld, Derek and Hoebers, Frank and
             Rietbergen, Michelle M. and Leemans, C. Rene and Dekker, Andre and
             Quackenbush, John and Gillies, Robert J. and Lambin, Philippe},
  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}
}

Downloads last month: 982