MedOtter/NSCLC-Radiogenomics

NSCLC-Radiogenomics

Non-small cell lung cancer (NSCLC) radiogenomic dataset on TCIA: pretreatment CT scans of 211 NSCLC patients with matching gene-expression, clinical, and mutation data. This HuggingFace mirror contains only the 144 patients with a DICOM SEG of the primary lung tumor (the segmentation-usable subset).

Dataset Details

PT/PET series (480 series, 201 patients) are excluded — the published segmentation masks are on CT only, so PET adds no signal for the segmentation task.

Annotation Pipeline

Per Bakr et al. (Sci. Data 2018), each segmentation was produced by:

1. 1.An unpublished automatic algorithm provided an initial mask of the primary tumor on the axial CT.
2. 2.Thoracic radiologist M.K. (5+ yrs experience) viewed every case and edited the masks as necessary in ePAD.
3. 3.Thoracic radiologist A.N.L. (20+ yrs experience) independently reviewed every case; disagreements were resolved by discussion between M.K. and A.N.L.
4. 4.Final approval by A.N.L.

Only one consolidated DICOM SEG is published per patient — it already reflects this consensus, so callers do not have to pick between annotators.

Cohorts (not splits)

The R01/AMC split is encoded in patient IDs (R01-xxx vs AMC-xxx). The SEG-having subset is not evenly distributed — see series_to_patient.json for the exact list of included patients. There is no predefined train/val/test split.

Structure

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

Each SEG references its source CT via ReferencedSeriesSequence (top-level CT SeriesInstanceUID) and PerFrameFunctionalGroupsSequence → DerivationImageSequence → SourceImageSequence (per-frame source CT SOPInstanceUID), enabling loss-less alignment to the CT slice grid.

Notes for Loaders

• —A patient may have multiple CT studies/series — pair the SEG to its exact referenced CT series, not the first CT under the patient ID.
• —DICOM SEG ⇄ ITK conversion is needed to get a labelmap volume; use dcmqi's segimage2itkimage or pydicom-seg.
• —All masks are binary (single primary-tumor foreground class).

Source

• —TCIA collection: https://www.cancerimagingarchive.net/collection/nsclc-radiogenomics/
• —DOI: 10.7937/K9/TCIA.2017.7hs46erv
• —Released: 2015 (Version 4, updated 2021-06-01). Fully public since 2025-07-07.

Citation

@article{bakr2018nsclcradiogenomics,
  author  = {Bakr, Shaimaa and Gevaert, Olivier and Echegaray, Sebastian and
             Ayers, Kelsey and Zhou, Mu and Shafiq, Majid and Zheng, Hong and
             Zhang, Weiruo and Leung, Ann and Kadoch, Michael and
             Shrager, Joseph and Quon, Andrew and Rubin, Daniel L. and
             Plevritis, Sylvia K. and Napel, Sandy},
  title   = {A radiogenomic dataset of non-small cell lung cancer},
  journal = {Scientific Data},
  volume  = {5},
  pages   = {180202},
  year    = {2018},
  doi     = {10.1038/sdata.2018.202}
}

@misc{nsclcradiogenomics2017tcia,
  author    = {Bakr, S. and Gevaert, O. and Echegaray, S. and Ayers, K. and
               Zhou, M. and Shafiq, M. and Zheng, H. and Zhang, W. and
               Leung, A. and Kadoch, M. and Shrager, J. and Quon, A. and
               Rubin, D. L. and Plevritis, S. K. and Napel, S.},
  title     = {Data for NSCLC Radiogenomics (Version 4) [Dataset]},
  year      = {2021},
  publisher = {The Cancer Imaging Archive},
  doi       = {10.7937/K9/TCIA.2017.7hs46erv}
}