PETWB-REP: A Multi-Cancer Whole-Body FDG PET/CT and Radiology Report Dataset for Medical Imaging Research

Le Xue, Gang Feng, Wenbo Zhang, Yichi Zhang, Lanlan Li, Shuqi Wang, Liling Peng, Sisi Peng, Xin GaoarXiv 2025 · DOI · arXiv

Introduction

In medical artificial intelligence, data availability remains a significant bottleneck. While deep learning has demonstrated remarkable potential in medical image analysis, existing public datasets are frequently "siloed," focusing on single organs or isolated imaging modalities. Positron Emission Tomography (PET) combined with Computed Tomography (CT) represents the clinical gold standard for cancer staging, as it integrates functional metabolic data with anatomical precision. However, large-scale PET/CT datasets paired with expert-written radiology reports—the "ground truth" of clinical interpretation—are exceedingly rare.

To bridge this gap, researchers have introduced PETWB-REP, a curated dataset featuring 490 whole-body FDG PET/CT scans paired with expert-validated radiology reports across a diverse range of malignancies. This multi-modal dataset provides a foundation for training AI models capable of automated tumor detection, segmentation, and clinical report generation.

Methodology and Data Preparation

The researchers implemented a rigorous pipeline to transform raw clinical data into a research-ready format. Data was retrospectively collected from 490 patients at the Shanghai Universal Medical Imaging Diagnostic Center between 2021 and 2024.

The preprocessing workflow involved several critical steps:

  • Normalization: Raw DICOM files were converted to NIfTI format. CT images underwent z-score normalization, while PET images were converted to Standardized Uptake Values (SUVbw) to ensure metabolic activity is comparable across different patients.
  • Spatial Alignment: To ensure perfect registration between modalities, B-spline interpolation was used to align PET images to the CT space.
  • Report Translation: Original Chinese radiology reports were translated into English using a two-step protocol: initial machine translation followed by meticulous manual review by a senior nuclear medicine physician to ensure terminology accuracy.
  • Anonymization: A double-check system was employed to strip all Protected Health Information (PHI) from both image headers and text reports.

The PETWB-REP Dataset

The dataset comprises 490 patients (219 female, 271 male) and covers over 20 types of cancer. Unlike highly controlled laboratory datasets, PETWB-REP reflects actual clinical referral patterns, providing a realistic distribution of cases encountered in a major imaging center.

Distribution of cancer types in the PETWB-REP dataset, highlighting the prevalence of lung, liver, and cervical cancers.

Key Dataset Features:

  • Paired Modalities: Every case includes CT, PET, and fused PET/CT volumes.
  • Bilingual Reports: De-identified radiology reports are provided in both English and Chinese.
  • Metadata: A comprehensive metadata.csv file includes clinical variables such as age, sex, and primary cancer type.

Vision-Language Integration

A core strength of PETWB-REP is the pairing of high-resolution functional imaging with descriptive clinical text. This enables research into "Multi-modal Fusion," where AI models learn to interpret visual metabolic signals and translate them into structured medical findings simultaneously.

Example of the paired data structure in PETWB-REP, showing whole-body PET/CT scans alongside their corresponding bilingual organ-specific reports.

Results and Technical Validation

The dataset shows a realistic clinical distribution, with lung cancer being the most prevalent (34.29%), followed by liver cancer (10.00%) and cervical cancer (7.76%). Technically, images were resampled to an isotropic in-plane resolution of 0.98 × 0.98 mm with a 3.00 mm slice thickness, ensuring high fidelity for granular lesion analysis. Every case underwent visual inspection by an experienced physician to confirm image quality and proper registration, ensuring the data is free from significant motion or metal artifacts.

Limitations and Future Directions

While PETWB-REP is a significant contribution, it faces certain limitations common in medical data:

  1. Single-Center Bias: The data reflects the specific equipment and reporting styles of one institution.
  2. Non-Uniform Distribution: Rare cancers have fewer samples, reflecting clinical reality but posing challenges for training models on infrequent pathologies.

Despite these hurdles, PETWB-REP is uniquely positioned for the development of medical "Foundation Models." These large-scale models can be fine-tuned for downstream tasks such as automated TNM staging, longitudinal treatment monitoring, and predicting patient outcomes based on complex radiomic features.

Summary updated

BibTeX
@misc{xue2025petwb,
  author = {Le Xue and Gang Feng and Wenbo Zhang and Yichi Zhang and Lanlan Li and Shuqi Wang and Liling Peng and Sisi Peng and Xin Gao},
  title = {{PETWB-REP: A Multi-Cancer Whole-Body FDG PET/CT and Radiology Report Dataset for Medical Imaging Research}},
  year = {2025},
  eprint = {2511.03194},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.03194},
}
Figure