MedOtter/Colorectal-Liver_Metastases

Colorectal-Liver-Metastases (CRLM)

Preoperative contrast-enhanced CT scans + manual radiologist segmentations from 197 patients who underwent hepatic resection for colorectal liver metastases at Memorial Sloan Kettering Cancer Center (MSKCC). Released under TCIA in 2023 alongside the Scientific Data descriptor by Simpson et al. (2024).

Dataset Details

Segment Categories (per patient, DICOM SEG / DSO)

Annotation provenance — Semi-automatic with Scout Liver (Pathfinder), then manually reviewed and corrected by an expert hepatobiliary radiologist or fellow at MSKCC. Single-annotator workflow; no inter-rater statistics are reported, so this is the sole and recommended ground truth.

Structure

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

Every SEG references its source CT via the standard DICOM SEG attributes:

• —top-level ReferencedSeriesSequence[0].SeriesInstanceUID (source CT series UID)
• —per-frame PerFrameFunctionalGroupsSequence → DerivationImageSequence → SourceImageSequence → ReferencedSOPInstanceUID (source CT slice UID per SEG frame)

so loaders can drop SEG frames onto the matching CT slice grid without resampling. series_to_patient.json mirrors the same metadata for fast indexing without re-reading DICOM headers.

Important Notes for Loaders

• —DICOM SEG ⇄ ITK conversion is required to obtain a labelmap volume. Use pydicom-seg or dcmqi's segimage2itkimage, or read the SEG pixel array and route each frame to its source CT slice (the loader utility under dataloader/colorectal_liver_metastases.py shows one reference implementation).
• —`Liver` overlaps tumors and vessels. The Liver segment is the whole-organ outline; voxel-level intersection with Tumor_*, Portal, and Hepatic masks is expected and must be resolved by the consumer (e.g. subtract tumor mask if you want non-tumor liver).
• —*`Tumor_ count varies per patient.** When merging into a single binary tumor mask, take the union of all Tumor_*` segments for that patient.
• —No predefined splits. The 197 cases are released as a single cohort; if you need a train/val/test partition, generate one yourself (e.g. by patient ID with a fixed seed).

Source

• —TCIA collection: https://www.cancerimagingarchive.net/collection/colorectal-liver-metastases/
• —DOI: 10.7937/QXK2-QG03 (Version 2, 2023)
• —Released: 2023, fully public since TCIA opened all collections (2025-07-07)

Citation

@article{simpson2024crlm,
  author  = {Simpson, Amber L. and Doussot, Alexandre and Creasy, John M. and
             Adams, Lauryn B. and Allen, Peter J. and DeMatteo, Ronald P. and
             Gönen, Mithat and Kemeny, Nancy E. and Kingham, T. Peter and
             Shia, Jinru and Jarnagin, William R. and Do, Richard K.G. and
             D'Angelica, Michael I.},
  title   = {Preoperative {CT} and survival data for patients undergoing
             resection of colorectal liver metastases},
  journal = {Scientific Data},
  volume  = {11},
  pages   = {172},
  year    = {2024},
  doi     = {10.1038/s41597-024-02981-2}
}

@misc{simpson2023crlm_tcia,
  author    = {Simpson, A. L. and others},
  title     = {Preoperative {CT} and Survival Data for Patients Undergoing
               Resection of Colorectal Liver Metastases
               (Colorectal-Liver-Metastases) (Version 2) [Dataset]},
  publisher = {The Cancer Imaging Archive},
  year      = {2023},
  doi       = {10.7937/QXK2-QG03}
}