PortPy_Dataset / croissant.jsonld
gourav3017's picture
Add Croissant metadata record (croissant.jsonld)
c824522 verified
Raw History Blame
26.9 kB
{
"@context": {
"@language": "en",
"@vocab": "https://schema.org/",
"arrayShape": "cr:arrayShape",
"citeAs": "cr:citeAs",
"column": "cr:column",
"conformsTo": "dct:conformsTo",
"containedIn": "cr:containedIn",
"cr": "http://mlcommons.org/croissant/",
"data": {
"@id": "cr:data",
"@type": "@json"
},
"dataBiases": "cr:dataBiases",
"dataCollection": "cr:dataCollection",
"dataType": {
"@id": "cr:dataType",
"@type": "@vocab"
},
"dct": "http://purl.org/dc/terms/",
"equivalentProperty": "cr:equivalentProperty",
"examples": {
"@id": "cr:examples",
"@type": "@json"
},
"extract": "cr:extract",
"field": "cr:field",
"fileProperty": "cr:fileProperty",
"fileObject": "cr:fileObject",
"fileSet": "cr:fileSet",
"format": "cr:format",
"includes": "cr:includes",
"isArray": "cr:isArray",
"isLiveDataset": "cr:isLiveDataset",
"jsonPath": "cr:jsonPath",
"key": "cr:key",
"md5": "cr:md5",
"parentField": "cr:parentField",
"path": "cr:path",
"personalSensitiveInformation": "cr:personalSensitiveInformation",
"recordSet": "cr:recordSet",
"references": "cr:references",
"regex": "cr:regex",
"repeated": "cr:repeated",
"replace": "cr:replace",
"samplingRate": "cr:samplingRate",
"sc": "https://schema.org/",
"separator": "cr:separator",
"source": "cr:source",
"subField": "cr:subField",
"transform": "cr:transform",
"rai": "http://mlcommons.org/croissant/RAI/",
"prov": "http://www.w3.org/ns/prov#"
},
"@type": "sc:Dataset",
"conformsTo": "http://mlcommons.org/croissant/1.1",
"name": "PortPy_Dataset",
"alternateName": [
"PortPy-Project/PortPy_Dataset",
"PortPy.AI benchmark dataset"
],
"description": "The PortPy.AI benchmark dataset is an AI-ready, optimization-ready radiotherapy treatment-planning dataset for 200 lung cancer and 129 prostate cancer patients, derived from the public, de-identified CT images and structure sets of two collections of The Cancer Imaging Archive (TCIA): NSCLC-Radiomics (lung) and Prostate-Anatomical-Edge-Cases (prostate). For every patient the dataset provides: the de-identified DICOM CT images and structure sets; PortPy-format CT and structure-mask arrays (HDF5) with metadata (JSON); physics-based beamlet-to-voxel dose-influence matrices for 72 candidate beams at 5-degree spacing (3 mm voxels, 2.5 mm beamlets restricted to the projected tumor plus a 5 mm margin), computed with the FDA-approved Eclipse treatment planning system (Anisotropic Analytical Algorithm) through its API; optimization-voxel data; expert-selected beam angles; harmonized structure names and clinical criteria (JSON); and deliverable reference treatment plans (IMRT, and VMAT where available; DICOM RTPLAN and RTDOSE) generated by the clinically deployed ECHO automated planning system (prioritized optimization) under a fixed protocol per disease site and reviewed by a board-certified medical physicist. The dataset supports three benchmark tasks - beamlet dose prediction, total dose prediction, and plan optimization - evaluated by both prediction accuracy and the clinical quality of the resulting plan, with frozen patient-level train/validation/test splits released in the PortPy toolkit (config_files/data_splits/data_splits_default.json). Total size is approximately 8 TB. Lightweight metadata is stored in JSON files and large arrays in HDF5 files; data_info.jsonl indexes the per-patient files. Toolkit and loaders: https://github.com/PortPy-Project/PortPy. Paper: PortPy.AI: A Benchmark for AI and Optimization in Cancer Radiotherapy Planning (NeurIPS 2026, Evaluations and Datasets Track).",
"url": "https://huggingface.co/datasets/PortPy-Project/PortPy_Dataset",
"license": "https://creativecommons.org/licenses/by-nc/4.0/",
"datePublished": "2026-05-03",
"version": "1.0.0",
"keywords": [
"radiotherapy",
"radiation therapy",
"treatment planning",
"dose prediction",
"beamlet dose prediction",
"dose-influence matrix",
"plan optimization",
"IMRT",
"VMAT",
"lung cancer",
"prostate cancer",
"medical physics",
"benchmark",
"cc-by-nc-4.0"
],
"citeAs": "@inproceedings{jhanwar2026portpyai,\n title={PortPy.AI: A Benchmark for AI and Optimization in Cancer Radiotherapy Planning},\n author={Jhanwar, Gourav and Tefagh, Mojtaba and Huang, Qijie and Golkarieh, Alireza and Hong, Linda and Pham, Hai and Taasti, Vicki Trier and Tuomaala, Seppo and Magliari, Anthony and Folkerts, Michael and Zhou, Ying and Yang, Jie and Nadeem, Saad and Zarepisheh, Masoud},\n booktitle={Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track},\n year={2026}\n}",
"creator": [
{"@type": "sc:Person", "name": "Gourav Jhanwar", "affiliation": {"@type": "sc:Organization", "name": "Memorial Sloan Kettering Cancer Center"}},
{"@type": "sc:Person", "name": "Mojtaba Tefagh", "affiliation": {"@type": "sc:Organization", "name": "University of Edinburgh"}},
{"@type": "sc:Person", "name": "Qijie Huang", "affiliation": {"@type": "sc:Organization", "name": "Memorial Sloan Kettering Cancer Center"}},
{"@type": "sc:Person", "name": "Alireza Golkarieh", "affiliation": {"@type": "sc:Organization", "name": "University of Michigan"}},
{"@type": "sc:Person", "name": "Linda Hong", "affiliation": {"@type": "sc:Organization", "name": "Memorial Sloan Kettering Cancer Center"}},
{"@type": "sc:Person", "name": "Hai Pham", "affiliation": {"@type": "sc:Organization", "name": "Memorial Sloan Kettering Cancer Center"}},
{"@type": "sc:Person", "name": "Vicki Trier Taasti", "affiliation": {"@type": "sc:Organization", "name": "Aarhus University Hospital"}},
{"@type": "sc:Person", "name": "Seppo Tuomaala", "affiliation": {"@type": "sc:Organization", "name": "Varian Medical Systems, Siemens Healthineers"}},
{"@type": "sc:Person", "name": "Anthony Magliari", "affiliation": {"@type": "sc:Organization", "name": "Varian Medical Systems, Siemens Healthineers"}},
{"@type": "sc:Person", "name": "Michael Folkerts", "affiliation": {"@type": "sc:Organization", "name": "Varian Medical Systems, Siemens Healthineers"}},
{"@type": "sc:Person", "name": "Ying Zhou", "affiliation": {"@type": "sc:Organization", "name": "Memorial Sloan Kettering Cancer Center"}},
{"@type": "sc:Person", "name": "Jie Yang", "affiliation": {"@type": "sc:Organization", "name": "Memorial Sloan Kettering Cancer Center"}},
{"@type": "sc:Person", "name": "Saad Nadeem", "affiliation": {"@type": "sc:Organization", "name": "Memorial Sloan Kettering Cancer Center"}},
{"@type": "sc:Person", "name": "Masoud Zarepisheh", "email": "zarepism@mskcc.org", "affiliation": {"@type": "sc:Organization", "name": "Memorial Sloan Kettering Cancer Center"}}
],
"publisher": {
"@type": "sc:Organization",
"name": "PortPy-Project",
"url": "https://huggingface.co/PortPy-Project"
},
"distribution": [
{
"@type": "cr:FileObject",
"@id": "repo",
"name": "repo",
"description": "The Hugging Face git repository (parquet conversion branch used for the machine-readable patient index).",
"contentUrl": "https://huggingface.co/datasets/PortPy-Project/PortPy_Dataset/tree/refs%2Fconvert%2Fparquet",
"encodingFormat": "git+https",
"sha256": "https://github.com/mlcommons/croissant/issues/80"
},
{
"@type": "cr:FileSet",
"@id": "parquet-files-for-config-default",
"description": "Parquet conversion of data_info.jsonl: one row per patient with the relative paths of that patient's files.",
"containedIn": {
"@id": "repo"
},
"encodingFormat": "application/x-parquet",
"includes": "default/*/*.parquet"
},
{
"@type": "cr:FileObject",
"@id": "repo-main",
"name": "repo-main",
"description": "The Hugging Face git repository (main branch) holding the per-patient data folders under data/<patient_id>/.",
"contentUrl": "https://huggingface.co/datasets/PortPy-Project/PortPy_Dataset/tree/main",
"encodingFormat": "git+https",
"sha256": "https://github.com/mlcommons/croissant/issues/80"
},
{
"@type": "cr:FileSet",
"@id": "patient-hdf5-arrays",
"description": "Per-patient HDF5 arrays: CT_Data.h5 (CT volume), StructureSet_Data.h5 (structure masks), and OptimizationVoxels_Data.h5 (optimization voxel data).",
"containedIn": {
"@id": "repo-main"
},
"encodingFormat": "application/x-hdf5",
"includes": "data/*/*.h5"
},
{
"@type": "cr:FileSet",
"@id": "patient-json-metadata",
"description": "Per-patient JSON metadata: CT_MetaData.json, StructureSet_MetaData.json (harmonized structure names), OptimizationVoxels_MetaData.json, and PlannerBeams.json (expert-selected beam angles).",
"containedIn": {
"@id": "repo-main"
},
"encodingFormat": "application/json",
"includes": "data/*/*.json"
},
{
"@type": "cr:FileSet",
"@id": "beam-hdf5-arrays",
"description": "Per-beam HDF5 files (data/<patient_id>/Beams/Beam_<k>_Data.h5) holding the beamlet-to-voxel dose-influence matrix and beamlet geometry of each candidate beam.",
"containedIn": {
"@id": "repo-main"
},
"encodingFormat": "application/x-hdf5",
"includes": "data/*/Beams/*.h5"
},
{
"@type": "cr:FileSet",
"@id": "beam-json-metadata",
"description": "Per-beam JSON metadata (data/<patient_id>/Beams/Beam_<k>_metadata.json): gantry angle, beamlet layout, and related beam parameters.",
"containedIn": {
"@id": "repo-main"
},
"encodingFormat": "application/json",
"includes": "data/*/Beams/*.json"
},
{
"@type": "cr:FileSet",
"@id": "dicom-files",
"description": "Per-patient DICOM files (data/<patient_id>/DicomFiles/): the de-identified CT image series and structure set (RTSTRUCT) as released by TCIA, and the reference-plan RTPLAN and RTDOSE objects generated by the ECHO automated planning system.",
"containedIn": {
"@id": "repo-main"
},
"encodingFormat": "application/dicom",
"includes": "data/*/DicomFiles/**/*.dcm"
}
],
"recordSet": [
{
"@type": "cr:RecordSet",
"dataType": "cr:Split",
"key": {
"@id": "default_splits/split_name"
},
"@id": "default_splits",
"name": "default_splits",
"description": "Splits for the default config. The index lists all patients in a single 'train' subset; the frozen benchmark train/validation/test partitions by patient ID are released with the PortPy toolkit in config_files/data_splits/data_splits_default.json.",
"field": [
{
"@type": "cr:Field",
"@id": "default_splits/split_name",
"dataType": "sc:Text"
}
],
"data": [
{
"default_splits/split_name": "train"
}
]
},
{
"@type": "cr:RecordSet",
"@id": "default",
"description": "Patient index (one record per patient, from data_info.jsonl): the patient identifier and the relative paths of that patient's data files within the repository.",
"field": [
{
"@type": "cr:Field",
"@id": "default/split",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"fileProperty": "fullpath"
},
"transform": {
"regex": "default/(?:partial-)?(train)/.+parquet$"
}
},
"references": {
"field": {
"@id": "default_splits/split_name"
}
}
},
{
"@type": "cr:Field",
"@id": "default/patient_id",
"description": "Patient identifier, e.g. Lung_Patient_1 or Prostate_Patient_1; also the name of the patient's data folder.",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "patient_id"
}
}
},
{
"@type": "cr:Field",
"@id": "default/ct_data",
"description": "Relative path to CT_Data.h5 (CT volume, HDF5).",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "ct_data"
}
}
},
{
"@type": "cr:Field",
"@id": "default/ct_metadata",
"description": "Relative path to CT_MetaData.json (CT geometry and acquisition metadata).",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "ct_metadata"
}
}
},
{
"@type": "cr:Field",
"@id": "default/structureset_data",
"description": "Relative path to StructureSet_Data.h5 (binary masks of the tumor targets and organs at risk).",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "structureset_data"
}
}
},
{
"@type": "cr:Field",
"@id": "default/structureset_metadata",
"description": "Relative path to StructureSet_MetaData.json (harmonized structure names and structure attributes).",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "structureset_metadata"
}
}
},
{
"@type": "cr:Field",
"@id": "default/beam_data_paths",
"description": "Relative paths to the per-beam HDF5 files (Beams/Beam_<k>_Data.h5) containing the beamlet-to-voxel dose-influence matrix of each candidate beam.",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "beam_data_paths"
}
},
"isArray": true,
"arrayShape": "-1"
},
{
"@type": "cr:Field",
"@id": "default/beam_metadata_paths",
"description": "Relative paths to the per-beam JSON metadata files (Beams/Beam_<k>_metadata.json).",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "beam_metadata_paths"
}
},
"isArray": true,
"arrayShape": "-1"
},
{
"@type": "cr:Field",
"@id": "default/optimization_voxels_data",
"description": "Relative path to OptimizationVoxels_Data.h5 (voxel coordinates and structure assignment used for optimization).",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "optimization_voxels_data"
}
}
},
{
"@type": "cr:Field",
"@id": "default/optimization_voxels_metadata",
"description": "Relative path to OptimizationVoxels_MetaData.json.",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "optimization_voxels_metadata"
}
}
},
{
"@type": "cr:Field",
"@id": "default/planner_beams",
"description": "Relative path to PlannerBeams.json (beam angles selected by an expert planner).",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "planner_beams"
}
}
},
{
"@type": "cr:Field",
"@id": "default/rt_dose_echo_imrt",
"description": "Relative path to the DICOM RTDOSE of the reference IMRT plan generated by the ECHO automated planning system.",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "rt_dose_echo_imrt"
}
}
},
{
"@type": "cr:Field",
"@id": "default/rt_plan_echo_imrt",
"description": "Relative path to the DICOM RTPLAN of the reference IMRT plan (deliverable beam shapes and intensities).",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "rt_plan_echo_imrt"
}
}
},
{
"@type": "cr:Field",
"@id": "default/rt_dose_echo_vmat",
"description": "Relative path to the DICOM RTDOSE of the reference VMAT plan, where available.",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "rt_dose_echo_vmat"
}
}
},
{
"@type": "cr:Field",
"@id": "default/rt_plan_echo_vmat",
"description": "Relative path to the DICOM RTPLAN of the reference VMAT plan, where available.",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "rt_plan_echo_vmat"
}
}
}
]
}
],
"rai:dataLimitations": "The release covers two disease sites, lung (200 patients) and prostate (129 patients), and therefore does not capture all anatomies, treatment protocols, and clinical scenarios encountered in radiotherapy practice. All data are for photon external-beam radiotherapy (IMRT and VMAT); other modalities such as proton therapy are not included. Beamlet dose-influence matrices are provided at a fixed discretization (72 candidate beams at 5-degree spacing, 3 mm voxels, 2.5 mm beamlets) and only for beamlets within the projected tumor plus a 5 mm margin, which is sufficient for treatment planning but not for arbitrary field geometries. Clinical criteria and reference plans follow the protocols of a single institution and were generated by a single automated planning system; they are intended as benchmark references rather than as universally optimal plans, and users can regenerate plans under other protocols with the released toolkit. The dataset and the accompanying baselines are intended solely for research and development and must not be used for clinical decision-making or patient treatment.",
"rai:dataBiases": "Patient selection is inherited from the two public source collections and may not represent the full diversity of cancer patients, disease stages, anatomies, imaging devices, or treatment practices. Patient demographics are not provided beyond what the source collections publish. Structure contours originate from the source collections and were harmonized in name but not re-contoured. Dose data reflect one treatment machine model and one commercial dose-calculation algorithm, and the reference plans reflect one institution's clinical criteria, beam-angle preferences, and automated planning system, so methods tuned to this dataset may inherit these institutional conventions.",
"rai:personalSensitiveInformation": "All patient data originate from two public, de-identified collections of The Cancer Imaging Archive (TCIA), which removes protected health information before release; patients are identified only by anonymized identifiers. The dataset redistributes the de-identified DICOM CT images and structure sets of these collections together with the derived planning data, in accordance with their licenses (NSCLC-Radiomics: CC BY-NC 3.0; Prostate-Anatomical-Edge-Cases: CC BY 4.0) and with attribution to TCIA and the collection authors. The lung and prostate scans do not include the head, so facial reconstruction and re-identification from the images are not possible. The dataset contains no protected health information. Users must comply with the TCIA data usage policies and with the CC BY-NC 4.0 license of this dataset, and must not attempt to re-identify patients.",
"rai:dataUseCases": "Research and benchmarking of AI and optimization methods for radiotherapy treatment planning, in particular the three PortPy.AI benchmark tasks: (1) beamlet-to-voxel dose prediction (learning a surrogate for the physics-based dose-influence matrix), (2) total dose prediction (predicting a patient-specific reference dose distribution from anatomy), and (3) plan optimization (computing deliverable beam shapes and intensities under clinical and machine constraints). Further intended uses include knowledge-based planning, automated planning, dose-influence-matrix compression, beam-angle optimization, multi-criteria optimization, and education. The dataset is not intended for clinical use, for training systems that make clinical decisions without expert oversight, or for any attempt to identify individuals.",
"rai:dataSocialImpact": "Positive impacts: the dataset lowers the barrier to radiotherapy planning research for groups without access to commercial treatment planning systems or clinical data, enables reproducible and comparable evaluation of AI and optimization methods, and can accelerate the development of faster and higher-quality automated planning, which ultimately benefits cancer patients. Potential negative impacts: research-grade models or plans derived from this dataset could be misused for clinical decisions without appropriate validation and regulatory approval, and methods tuned to one institution's protocols could generalize poorly elsewhere. Mitigations: the dataset and toolkit are released under a non-commercial license with an explicit research-only statement, reference plans are documented as benchmark references rather than clinical standards, and the evaluation protocol reports per-criterion clinical scorecards rather than a single score.",
"rai:hasSyntheticData": false,
"prov:wasDerivedFrom": [
{
"@id": "https://doi.org/10.7937/K9/TCIA.2015.PF0M9REI",
"prov:label": "NSCLC-Radiomics (The Cancer Imaging Archive), Aerts et al., version 4",
"sc:license": "CC BY-NC 3.0"
},
{
"@id": "https://doi.org/10.7937/QSTF-ST65",
"prov:label": "Prostate-Anatomical-Edge-Cases (The Cancer Imaging Archive), Kanwar et al.",
"sc:license": "CC BY 4.0"
}
],
"prov:wasGeneratedBy": [
{
"@type": "prov:Activity",
"prov:type": {
"@id": "https://www.wikidata.org/wiki/Q4929239"
},
"prov:label": "Data collection and derivation",
"sc:description": "The de-identified DICOM CT images and structure sets of 200 lung cancer patients (NSCLC-Radiomics) and 129 prostate cancer patients (Prostate-Anatomical-Edge-Cases) were downloaded from The Cancer Imaging Archive and imported into the FDA-approved Eclipse treatment planning system (Varian Medical Systems) at Memorial Sloan Kettering Cancer Center, in collaboration with the vendor. For each patient, a full 360-degree arc was discretized into 72 candidate beams at 5-degree spacing, the anatomy into 3 mm voxels, and each beam into 2.5 mm beamlets within the projected tumor plus a 5 mm margin; the dose deposited by every beamlet in every voxel was computed with the Anisotropic Analytical Algorithm through the Eclipse API, taking days of computation per patient, and assembled into the beamlet-to-voxel dose-influence matrices. No clinical patient data from the institution were used."
},
{
"@type": "prov:Activity",
"prov:type": {
"@id": "https://www.wikidata.org/wiki/Q5227332"
},
"prov:label": "Data preprocessing",
"sc:description": "The derived data were converted into the PortPy data format: HDF5 files for large arrays (CT volume, structure masks, optimization voxels, per-beam dose-influence matrices) and JSON files for metadata. Structure names were harmonized across patients under expert guidance so that each clinical criterion can be linked to a consistently named structure. Clinical criteria for the lung and prostate protocols were curated from current clinical practice and released as JSON files. A per-patient index (data_info.jsonl) lists the files of each patient."
},
{
"@type": "prov:Activity",
"prov:type": {
"@id": "https://www.wikidata.org/wiki/Q109719325"
},
"prov:label": "Reference plan generation and expert review",
"sc:description": "Beam angles were selected by an expert planner. Reference treatment plans were generated with the ECHO automated planning system (prioritized optimization with deliverability enforced by column generation and sequential convex programming), which is integrated into the institution's routine clinical workflow, under a fixed clinical protocol per disease site, and were reviewed by a board-certified medical physicist. Because generation is algorithmic and protocol-fixed, planner-to-planner variability is largely removed. The plans are exported as DICOM RTPLAN and RTDOSE objects."
},
{
"@type": "prov:Activity",
"prov:type": {
"@id": "https://www.wikidata.org/wiki/Q3306762"
},
"prov:label": "Quality control",
"sc:description": "Patient folders and file names follow a consistent convention across the dataset, and all files are loaded and checked with the PortPy toolkit, whose data loaders and tutorial notebooks exercise each data component. Frozen patient-level train/validation/test splits (data_splits_default.json) are released with the toolkit so that benchmark results are comparable across groups; all beamlets and voxels of a patient belong to a single partition."
}
]
}