This End User Data Use Agreement for Academic and Institutional Users (this
“Data Use Agreement”) is by and between you and University of Wisconsin-
Madison (the “University”) and governs your access to and use of the deidentified
imaging and associated demographic and clinical data provided to you by the
University through Hugging Face (“University Data”). By accessing, using, or
downloading any University Data, you are agreeing to be bound by the terms and
conditions of this Agreement.

The University is the legal entity that owns the University Data.
The University grants you a limited, personal, worldwide, non-exclusive, non-
transferable, royalty-free license to view and use the University Data solely for
your personal, Non-Commercial Research Purposes. Any commercial use, sale, or
other monetization is prohibited. “Non-Commercial Research Purposes” means
academic research or other not-for-profit scholarly purposes which are
undertaken at a non-profit or government institution and publishing in
connection therewith.
Other than the rights granted herein, the University retains all rights, title, and
interest in the University Data.
You will use the University Data only in compliance with all applicable laws
and regulations.
You may make a verbatim copy of the University Data for personal, non-
commercial research use. If another user within your organization wishes to use
the University Data, they must separately download the data set and enter into a
Data Use Agreement directly with the University.
You may not distribute, publish, or reproduce a copy of any portion or all of the
University Data to others without specific prior written permission from the
University.
You will promptly report to the University in writing any unauthorized access
or use of the University Data.
You will register with and provide current contact information to Hugging
Face. By accessing, using, or downloading the University Data, you consent to be
contacted by the University. The University may use, and you hereby grant the
University a non-exclusive, worldwide, perpetual, irrevocable license to use, any
data you provide to the University, including any personally identifiable
information.
You will acknowledge your use of University Data in all oral or written
presentations, disclosures, or publications arising from your use of the University
Data.
You may not modify, reverse engineer, decompile, or create derivative works
from the University Data. For avoidance of doubt, users may create research
results, including any data analyses, results of algorithm validations, annotations
and overlays, and outcomes or evaluations using the University Data. The
University will have no right, title, or interest in any such research results. You
must not remove or alter any copyright or other proprietary notices in the
University Data. You are encouraged to submit any annotations or overlays
created using the University data.
The University Data has not been reviewed or approved by the Food and Drug
Administration, and is for non-clinical, research use only. In no event shall data
or images generated through the use of the University Data be used or relied
upon in the diagnosis or provision of patient care.
THE UNIVERSITY DATA IS PROVIDED ON AN ”AS IS”, “WHERE IS” BASIS. THE
UNIVERSITY MAKES NO WARRANTIES, EXPRESS OR IMPLIED, INCLUDING BUT
NOT LIMITED TO WARRANTIES OF MERCHANTABILITY, FITNESS FOR A
PARTICULAR PURPOSE, AND NON INFRINGEMENT. THE UNIVERSITY WILL HAVE
NO LIABILITY WHATSOEVER ARISING FROM YOUR ACCESS OR USE OF THE
UNIVERSITY DATA.
You will not make any attempt to re-identify any individual from the
University Data. In the event that you identify any person whose data is included
in the University Data, you will promptly delete the data and notify the University
in writing.
The University may terminate this Data Use Agreement at any time upon
notice to you. The University shall have no liability in the event of any such
termination. Upon such termination, the University may direct that you
immediately return or destroy all copies of any University Data and, upon the
University’s request, you will certify the foregoing to the University in writing.

PETARSeg-11K

PETARSeg-11K is a large-scale PET/CT dataset designed for lesion-level, spatially grounded vision-language research. It was introduced in PETAR: Localized Findings Generation with Mask-Aware Vision-Language Modeling for PET Automated Reporting.

PETARSeg-11K was developed to provide explicit correspondence between localized abnormalities in PET/CT imaging and the free-text radiological findings describing those abnormalities. This lesion-level correspondence supports research in areas including localized findings generation, PET/CT visual grounding, lesion-level image-text alignment, mask-guided vision-language modeling, and automated PET reporting.

Citation

If you use PETARseg-11k, please cite these two papers:

@InProceedings{Maqbool_2026_CVPR,
    author    = {Maqbool, Danyal and Lee, Changhee and Huemann, Zachary and Church, Samuel D. and Larson, Matthew E. and Perlman, Scott B. and Romero, Tomas A. and Warner, Joshua D. and Lubner, Meghan and Tie, Xin and Merkow, Jameson and Hu, Junjie and Cho, Steve Y. and Bradshaw, Tyler J.},
    title     = {PETAR: Localized Findings Generation with Mask-Aware Vision-Language Modeling for PET Automated Reporting},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    year      = {2026},
    pages     = {42637-42648},
    shorttitle = {{PETAR}},
    url = {http://arxiv.org/abs/2510.27680},
    doi = {10.48550/arXiv.2510.27680},
    note = {arXiv:2510.27680 [cs]}
}

@article{huemann_contextual_2026,
    title = {{ConTEXTual} {Net} {3D}: {Vision}-{Language} {Modeling} in {PET}/{CT} for {Visual} {Grounding} of {Positive} {Findings}},
    issn = {2948-2933},
    shorttitle = {{ConTEXTual} {Net} {3D}},
    url = {https://doi.org/10.1007/s10278-026-01879-2},
    doi = {10.1007/s10278-026-01879-2},
    language = {en},
    urldate = {2026-07-27},
    journal = {Journal of Imaging Informatics in Medicine},
    author = {Huemann, Zachary and Church, Samuel and Warner, Joshua D. and Tran, Daniel and Tie, Xin and McMillan, Alan B. and Hu, Junjie and Cho, Steve Y. and Lubner, Meghan G. and Bradshaw, Tyler J.},
    month = apr,
    year = {2026},
}

Dataset Organization

This Hugging Face release provides the PET/CT imaging data in their DICOM representation.

Each anonymized patient/study is stored in its own directory in a tarball. The directory contains the corresponding CT DICOM series, PET DICOM series, and a DICOM RT Structure Set (RTSTRUCT) containing the lesion segmentations for that examination.

A patient directory is organized as follows:

PETWB_XXXXXX_XX/
├── CT/
│   └── <CT series>/
│       ├── *.dcm
│       ├── *.dcm
│       └── ...
└── PT/
    ├── <PET series>/
    │   ├── *.dcm
    │   ├── *.dcm
    │   └── ...
    └── patient_RTSTRUCT.dcm

The major components are:

CT/ — contains the CT DICOM series for the examination.
PT/ — contains the PET DICOM series for the examination.
PT/patient_RTSTRUCT.dcm — contains the lesion segmentation contours associated with the examination.

The RTSTRUCT contains the segmentation information for the annotated lesions in a study. Therefore, separate segmentation image files are not required for each lesion as the lesion annotations for a patient are stored together within the corresponding RTSTRUCT.

Segmentation-to-Text Annotations

PETARSeg-11K pairs each localized lesion segmentation with a corresponding free-text radiological finding.

The segmentation-to-text annotations are provided in JSON format. The top-level keys are anonymized patient/study identifiers. Within each patient, individual lesion annotations are identified as mask1, mask2, mask3, and so on.

For example:

{
  "PETWB_009794_01": {
    "mask1": "A focus of increased metabolic activity ...",
    "mask2": "A new hypermetabolic focus ...",
    "mask3": "The largest and most hypermetabolic ..."
  },
  "PETWB_008513_01": {
    "mask1": "Left level IIa lymph node ..."
  }
}

For each patient:

The patient ID corresponds to the anonymized patient/study directory containing the PET, CT, and RTSTRUCT data.
Each maskN entry corresponds to an individual lesion segmentation for that patient.
The value of each maskN entry is the free-text radiological description associated with that lesion.
Patients may contain one or multiple lesion annotations depending on the number of localized findings associated with the examination.

Metadata

The metadata/ directory contains CSV files describing the train and test splits:

metadata/
├── train.csv
└── test.csv

Each row corresponds to an anonymized PET/CT examination and provides paths to the associated imaging data.

The metadata contains the following fields:

Anonymised patient ID — anonymized identifier for the patient/study.
PT Dicom path — path to the PET DICOM series directory.
CT Dicom path — path to the CT DICOM series directory.
RTStruct path — path to the corresponding patient_RTSTRUCT.dcm file.

An example entry is:

Anonymised patient ID: PETWB_XXXXXX_XX
PT Dicom path: testing/PETWB_XXXXXX_XX/PT/<PET series>
CT Dicom path: testing/PETWB_XXXXXX_XX/CT/<CT series>
RTStruct path: testing/PETWB_XXXXXX_XX/PT/patient_RTSTRUCT.dcm

Relationship Between the Files

For a given patient, the dataset can be interpreted as:

PETWB_XXXXXX_XX
│
├── PET DICOM series  ─────────────┐
│                                  │
├── CT DICOM series   ─────────────┼── Spatial imaging context
│                                  │
├── patient_RTSTRUCT.dcm ──────────┘
│            │
│            ├── lesion / mask 1
│            ├── lesion / mask 2
│            └── lesion / mask N
│
└── JSON annotations
             │
             ├── mask1 → radiological finding for lesion 1
             ├── mask2 → radiological finding for lesion 2
             └── maskN → radiological finding for lesion N

Dataset Background

Additional details regarding dataset construction, annotation generation, and the PETAR-4B model are available in the PETAR paper and ConTEXTual Net 3D paper.

Intended Use

PETARSeg-11K is intended for non-commercial research use in accordance with the Data Use Agreement associated with this dataset.

The dataset is intended for research purposes only and is not authorized for direct clinical decision-making or patient care.

If you wish to use this dataset for commercial purposes, please reach out to tbradshaw@wisc.edu.

Downloads last month: 278

Total file size:

327 GB

Paper for UW-Madison-Dept-Radiology/PETARSeg-11K

PETAR: Localized Findings Generation with Mask-Aware Vision-Language Modeling for PET Automated Reporting

Paper • 2510.27680 • Published Nov 30, 2025