MOOSE and TotalSegmentator anatomy segmentations and radiomics features for CPTAC and CMB CT images in NCI Imaging Data Commons

This dataset contains volumetric segmentations of anatomical structures, generated automatically by AI models, and first-order and shape radiomics features for each segmented structure, for CT images available from the NCI Imaging Data Commons (IDC) [1]. This version covers CT series from 18 collections in IDC release v24, from two NCI programs: 9 collections of the Clinical Proteomic Tumor Analysis Consortium (CPTAC) and 9 of the Cancer Moonshot Biobank (CMB). Both programs published their images through The Cancer Imaging Archive (TCIA), and IDC hosts them in DICOM; each collection and its source DOI is listed in the table below.

These segmentations and radiomics features will be deposited to IDC in a future data release, where they will be available alongside the CT images they were derived from.

Cohort

The CT series were selected with the query used for the TotalSegmentator-CT-Segmentations dataset [2], applied to IDC v24 and restricted to the CPTAC and CMB collections: 3,059 CT series from 1,113 studies of 578 patients. This dataset has results for 3,035 of those series (1,102 studies, 574 patients).

Models

Every series was processed with two model families:

• MOOSE v3.2.2 [3] (https://github.com/ENHANCE-PET/MOOSE): ten clinical CT models — clin_ct_body, clin_ct_body_composition, clin_ct_cardiac, clin_ct_digestive, clin_ct_lungs, clin_ct_muscles, clin_ct_organs, clin_ct_peripheral_bones, clin_ct_ribs and clin_ct_vertebrae
• TotalSegmentator v2.18.0 [4] (https://github.com/wasserth/TotalSegmentator): the total and lung_vessels tasks

For each series and model there is one DICOM Segmentation (SEG) object holding that model's structures as binary segments, labelled with SNOMED CT codes, and one DICOM Structured Report (SR) following template TID 1500, holding the first-order and shape features of each segment computed with Radiomics.jl v2.0.0. SEG and SR objects were encoded with dcmqi [5]. Radiomics.jl skips segments larger than 5 million voxels.

In total there are 35,625 SEG and 35,073 SR objects (70,698 files, 14.8 GiB uncompressed). The analysis ran on the Terra platform with the CloudSegmentator workflow (https://github.com/ImagingDataCommons/CloudSegmentator) [6].

Every SEG and SR object records the DOI of this dataset in Other Clinical Trial Protocol IDs Sequence (0012,0023), with Issuer of Clinical Trial Protocol ID "DOI", and its source collection in Clinical Trial Sponsor Name (0012,0010) and Clinical Trial Protocol ID (0012,0020).

Files

There is one zip file per collection, named by its IDC collection_id (for example cptac_luad.zip):

Inside each zip, every SEG and SR object is in a folder of its own:

<collection_id>/<PatientID>/<StudyInstanceUID>/<CT SeriesInstanceUID>_SEG_<SEG SeriesInstanceUID>/<model>_0.dcm
<collection_id>/<PatientID>/<StudyInstanceUID>/<CT SeriesInstanceUID>_SR_<SR SeriesInstanceUID>/<model>_0_sr.dcm

The folder name starts with the SeriesInstanceUID of the CT series that was segmented, which is what you need to download the images, and ends with the SeriesInstanceUID of the SEG or SR series itself. For example, the MOOSE clin_ct_ribs segmentation of one CT series of patient C3N-00199 is

cptac_luad/C3N-00199/1.3.6.1.4.1.14519.5.2.1.7085.2626.183894862458344908946921283818/1.3.6.1.4.1.14519.5.2.1.7085.2626.386536609837235111399493634163_SEG_1.2.276.0.7230010.3.1.3.296485376.3207.1792152747.53782/clin_ct_ribs_0.dcm

A CT series has up to 12 SEG and 12 SR folders, one of each per model. Viewers tell them apart by Series Description, for example MOOSE(v3.2.2) clin_ct_ribs Segmentation, TotalSegmentator(v2.18.0) total Segmentation or MOOSE v3.2.2 clin_ct_ribs Radiomics.

Getting the CT images

The CT images are not included here. They are public in IDC and can be downloaded without an account using idc-index (pip install --upgrade idc-index). Run the download from the folder where you extracted the zip: idc-index saves each series to <collection_id>/<PatientID>/<StudyInstanceUID>/CT_<SeriesInstanceUID>/, so the images land in the same study folder as their segmentations.

To download one series, give its CT SeriesInstanceUID, here the one from the example above:

idc download 1.3.6.1.4.1.14519.5.2.1.7085.2626.386536609837235111399493634163

To download every CT series segmented in one collection, in Python:

from pathlib import Path
from idc_index import IDCClient

ct_uids = sorted({p.name.split("_")[0] for p in Path("cptac_luad").glob("*/*/*_SEG_*")})
IDCClient().download_from_selection(seriesInstanceUID=ct_uids, downloadDir=".", dry_run=True)

The images are far larger than the segmentations: the 165 CT series referenced in cptac_luad.zip (0.3 GB) take 26 GB. With dry_run=True the call reports the size and downloads nothing; remove it to download. The same series can also be downloaded inside 3D Slicer with the IDCBrowser extension.

Viewing the images and segmentations

Viewers that support DICOM Segmentation overlay each SEG on the CT series it was derived from. Import the CT and its segmentations together; after the download described above, the study folder holds both.

• 3D Slicer 5.12 and later read DICOM SEG and TID 1500 SR as installed; earlier versions need the QuantitativeReporting extension, from the Extensions Manager. Import the folder in the DICOM module, then select the CT and SEG series and load them. Segments appear in the Segmentations module with their names and SNOMED CT codes, and an SR loads as a table of its radiomics features.
• OHIF Viewer runs in the web browser and reads images from a DICOMweb server. Import the CT and SEG files into one, for example a Google Cloud DICOM store, and open the study in an OHIF instance connected to that server; selecting a SEG series overlays it on the CT. For the radiomics features in the SR, use 3D Slicer or the tools below.
• Other DICOM software that supports DICOM Segmentation, including PACS, can import the files the same way.

The SEG objects use the Deflated Explicit VR Little Endian transfer syntax (1.2.840.10008.1.2.1.99), which keeps them small but which some older DICOM software cannot read. DCMTK rewrites such a file uncompressed with dcmconv +te in.dcm out.dcm. Expect a much larger file: in one test, a 57 KB SEG became 18 MB.

To work with the files in code, pydicom reads both object types, and highdicom provides classes for SEG and SR (pip install pydicom highdicom). The dcmqi [5] command-line tools convert a SEG to NIfTI or NRRD (segimage2itkimage) and an SR to JSON (tid1500reader); pip install dcmqi installs them on Linux, macOS and Windows.

Known limitations

• The segmentations were generated automatically and were not reviewed by experts.
• Of the 36,708 possible results (3,059 series × 12 models), 1,083 are absent. 306 were lost to processing failures, 554 because the model found none of its structures in the image (for example, no lungs in a pelvic CT), and 223 because clin_ct_body_composition produced no output. 24 series have no results at all.
• 552 SEG objects have no corresponding SR (543 MOOSE, 537 of them clin_ct_body; 9 TotalSegmentator), mostly because every segment exceeded the radiomics size limit.
• 10 SEG objects lack Referenced Series Sequence (0008,1115): the MOOSE clin_ct_digestive, clin_ct_lungs, clin_ct_muscles, clin_ct_ribs and clin_ct_vertebrae results for two series of patient MSB-08964 (cmb_lca). The segmented CT series is still identified by the folder name, but software that looks for it in that attribute will not find it.

Citation

If you use this dataset, please cite it, and the publications describing the models and the IDC.

Support

For questions about this dataset, or to report issues, contact Imaging Data Commons support at support@canceridc.dev or (preferred) the IDC Forum at https://discourse.canceridc.dev.

Acknowledgments

This work was supported in part by the NCI Imaging Data Commons and The Cancer Imaging Archive projects, funded in whole or in part with Federal funds from the National Cancer Institute, National Institutes of Health, Department of Health and Human Services, under Task Order Nos. HHSN26110071 and 75N91019F00131, under Contract No. HHSN261201500003I. The content of this publication does not necessarily reflect the views or policies of the Department of Health and Human Services, nor does mention of trade names, commercial products, or organizations imply endorsement by the U.S. Government.

This work was also supported by NIH grant U24CA306986.

References

1. Fedorov A, et al. National Cancer Institute Imaging Data Commons: Toward Transparency, Reproducibility, and Scalability in Imaging Artificial Intelligence. RadioGraphics (2023). https://doi.org/10.1148/rg.230180
2. Thiriveedhi VK, Krishnaswamy D, Clunie D, Fedorov A. TotalSegmentator-CT-Segmentations: TotalSegmentator segmentations and radiomics features for NCI Imaging Data Commons CT images. Zenodo. https://doi.org/10.5281/zenodo.8347011
3. Shiyam Sundar LK, et al. Fully Automated, Semantic Segmentation of Whole-Body 18F-FDG PET/CT Images Based on Data-Centric Artificial Intelligence. J Nucl Med (2022). https://doi.org/10.2967/jnumed.122.264063
4. Wasserthal J, et al. TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images. Radiol Artif Intell (2023). https://doi.org/10.1148/ryai.230024
5. Herz C, et al. dcmqi: An Open Source Library for Standardized Communication of Quantitative Image Analysis Results Using DICOM. Cancer Research (2017). https://doi.org/10.1158/0008-5472.CAN-17-0336
6. Thiriveedhi VK, et al. Cloud-based large-scale curation of medical imaging data using AI segmentation. Research Square preprint (2024). https://doi.org/10.21203/rs.3.rs-4351526/v1