PET Dataset: Modalities & Applications

• PET dataset is a term that encompasses diverse resources including PET/CT imaging, Pascal eye-tracking, and process extraction from text, each with distinct modalities.
• Applications span image quality assessment, lesion segmentation, gaze analysis, and business process extraction, highlighting task-specific design.
• Implications include the need for precise disambiguation and tailored processing pipelines to handle varied annotations, formats, and access conditions.

Searching arXiv for the cited PET-related datasets and the PET eye-tracking dataset to ground the article in current records. The expression “PET dataset” does not denote a single resource. In arXiv literature, it refers to multiple datasets across distinct domains: positron emission tomography and PET/CT imaging, the Pascal animal classes Eye-Tracking database, and a corpus for Process Extraction from Text. Within medical imaging alone, the label covers image-quality assessment datasets, whole-body FDG PET/CT and radiology report corpora, organ and lesion segmentation resources, slice-classification cohorts, and synthetic or translated PET data for reconstruction and image-to-image generation (Li et al., 25 Jun 2025, Xue et al., 5 Nov 2025, Nguyen et al., 29 Sep 2025, Gilani et al., 2016, Bellan et al., 2022). Because the acronym is overloaded, precise usage requires explicit specification of modality, task, annotation protocol, and access conditions.

1. Terminological scope and disambiguation

In medical imaging, “PET” ordinarily denotes Positron Emission Tomography, often in PET/CT settings. In computer vision, however, PET may denote the Pascal animal classes Eye-Tracking dataset, and in natural language processing it may denote the Process Extraction from Text corpus. This ambiguity is not merely terminological; it changes the unit of analysis from volumetric scans and SUV maps to gaze trajectories or token-span annotations.

Dataset name	Domain	Core contents
PET-CT-IQA-DS	PET/CT image quality assessment	2,700 varying-quality PET/CT images with radiologist quality scores
PETWB-REP	Whole-body FDG PET/CT + reports	490 patients with PET, CT, reports, and metadata
ViMed-PET	PET/CT + Vietnamese reports	2,757 studies, 1,567,062 paired slices, 2,757 reports
PETS-5k	Whole-body PET segmentation	5,731 3D PET volumes with multi-organ labels
PET	Eye tracking	4,135 PASCAL VOC animal images with gaze recordings
PET	Process extraction from text	45 annotated process descriptions

A common misconception is that “PET dataset” identifies a canonical benchmark. The literature summarized here indicates the opposite: PET datasets are task-specific, with incompatible annotations, different spatial units, and substantially different release models ranging from open download to controlled access or institution-held collections.

2. PET/CT image quality assessment datasets

The dataset PET-CT-IQA-DS was introduced to “fill the blank of PET/CT IQA dataset” and accompanies the MS-IQA framework for PET/CT image quality assessment (Li et al., 25 Jun 2025). It contains 2,700 simulated PET/CT image pairs, described as 2,700 distorted PET/CT slices. The pristine base set comprises 20 patients × 5 anatomically distinct slices per patient = 100 “clean” PET/CT pairs drawn from the Lung-PET-CT-Dx public repository. Three distortion sources are applied—Poisson noise, Gaussian noise and JPEG compression—each at three severity levels, yielding 27 distortion combinations per clean image and hence 27 × 100 = 2,700 total samples.

Each sample carries a single continuous quality score in $[0,4]$, with higher = better. Annotation was performed by five board-certified radiologists using a single-image, no-reference subjective IQA protocol evaluating visual distortion and diagnostic utility. The scoring scale is a 5-point Likert from 0 (“unacceptable”) to 4 (“excellent”). The ground-truth score is the arithmetic mean of the five ratings:

$Q_i = \frac{1}{5}\sum_{r=1}^5 q_{i,r},$

where $q_{i,r}\in\{0,1,2,3,4\}$.

The publication is explicit about several omissions. Source scanner models are not reported; patient demographics are not reported beyond “20 patients”; spatial resolution / voxel size / slice thickness are not specified; tracer injection activity / timing are not specified; preprocessing is not described aside from the synthetic distortion pipeline; inter-rater agreement is not reported; and histogram or bin counts of the $Q_i$ values are not provided. This suggests that PET-CT-IQA-DS is primarily a controlled distortion-and-rating resource rather than a fully characterized acquisition dataset.

For downstream usage, the authors employ an 80%/20% train/test partition on a patient-wise basis to avoid leakage. That design choice aligns the dataset with regression-style IQA evaluation, since all downstream experiments use the MOS in $[0,4]$ as the regression target.

3. Whole-body PET/CT and radiology report corpora

The dataset PETWB-REP is a curated whole-body $^{18}$F-FDG PET/CT and radiology report dataset comprising 490 patients, including 219 F, 271 M; mean age 60.98 ± 12.77 years, with cancer types including lung cancer, liver cancer, breast cancer, prostate cancer, and ovarian cancer (Xue et al., 5 Nov 2025). It provides 490 whole-body $^{18}$F-FDG PET series and 490 low-dose CT series. Acquisition was performed on a Siemens Biograph 64 PET/CT scanner, with $\ge 6$ h fasting, blood glucose < 11.1 mmol/L, $^{18}$F-FDG, 3.70–5.55 MBq/kg IV, and 60 min rest in a dim room. PET acquisition used 3D mode, 5–6 bed positions, 2.5 min per bed position, with OSEM (2 iterations, 21 subsets) + 5 mm Gaussian smoothing. Post-processing includes resampling to isotropic in-plane 0.98×0.98 mm, slice thickness 3.00 mm, and PET→CT co-registration via B-spline interpolation.

PETWB-REP integrates imaging and text at several levels. The reports are provided in both Chinese and English CSV, with sections by Head/Neck; Chest; Abdomen/Pelvis; Musculoskeletal, and include CT findings, PET findings, lesion entries with single transverse diameter, site, SUVmax, and an “Impression” summary. Structured metadata in meta_data.csv include age, sex, unique subject ID, primary cancer type, staging if available, treatment history (where recorded), and imaging-protocol fields such as injected dose, uptake time, scanner model, acquisition dates (de-identified). Data are organized as NIfTI (.nii.gz) plus anonymized original DICOM headers, with UTF-8 CSV for reports and metadata. Access is via Zenodo, under CC BY 4.0, with free download after Zenodo registration; no additional IRB needed.

The dataset ViMed-PET addresses a different axis of scarcity: PET/CT corpora paired with reports in a low-resource language (Nguyen et al., 29 Sep 2025). It contains 2,757 PET/CT “studies”, 1,567,062 paired axial slices of PET and CT, and 2,757 full-length Vietnamese clinical reports. Reports cover whole-body oncology workflows, including staging and monitoring of lung cancer, thyroid cancer, metastatic disease. Imaging was acquired on GE Discovery 710 PET/CT and GE Discovery STE PET/CT scanners with $^{18}$F-FDG tracer injected; PET volumes attenuation-corrected using the aligned CT; and coverage from head to upper thighs. The released raw dataset retains original voxel dimensions, with no further intensity normalization beyond the vendor’s SUV calibration.

ViMed-PET uses a structured report representation. Reports parsed from DOCX are converted to JSON and contain fixed sections including patient demographics; clinical history; indication; scanning protocol; region-by-region observations; impression. Fields include age, sex, weight, injected activity, uptake time, tracer type, and SUV metrics. Additional DICOM tags include patient age, sex, weight, height, radiotracer activity, scanner model, slice thickness, acquisition date. The dataset also defines multiple derived subsets: 8,271 region-specific image–report samples produced by partitioning each whole-body study into head–neck, chest, abdomen–pelvis segments with 20-slice overlap; a training split of 5,571 image–report pairs, validation of 975 pairs, test of 1,725 pairs; an expert-validated lung cancer “medical test set” of 80 cases with 398 lesions; a VQA subset of 8,271 conversations; and a study-comparison subset of 10,000 image–report comparison tuples. Availability is through a controlled-access research agreement.

Taken together, PETWB-REP and ViMed-PET illustrate two different multimodal design strategies. PETWB-REP emphasizes whole-body FDG PET/CT, reports, and structured clinical metadata in an openly downloadable resource, whereas ViMed-PET emphasizes PET/CT-report pairs in Vietnamese, augmented subsets, and expert-validated downstream benchmarks.

4. Segmentation, detection, and slice-level classification resources

The dataset PETS-5k was introduced for universal promptable segmentation from PET images and is described as the largest PET segmentation dataset to date (Zhang et al., 20 Feb 2025). It comprises 5,731 three-dimensional whole-body PET images and approximately 1.3 million 2D slices. Acquisition used $Q_i = \frac{1}{5}\sum_{r=1}^5 q_{i,r},$0F-FDG; injection dose 3.7 MBq/kg; uptake time median 67 min (range 53–81 min); $Q_i = \frac{1}{5}\sum_{r=1}^5 q_{i,r},$1 h fasting; blood glucose < 200 mg/dL; scanner Siemens Biograph mCT (PET/CT). CT acquisition used 120 kVp, 40–100 mAs, 5 mm slice thickness, and PET acquisition used 3D Flowmotion, 5 min per bed position, reconstructed with the TrueX algorithm at 4.07×4.07×3 mm$Q_i = \frac{1}{5}\sum_{r=1}^5 q_{i,r},$2 voxel size.

PETS-5k distinguishes between training targets and unseen targets. The five training organs are liver, left kidney, right kidney, heart, spleen. The seven unseen test-only organs are aorta, prostate, left lung lower lobe, right lung lower lobe, left lung upper lobe, right lung upper lobe, right lung middle lobe. The split is 5,631 low-quality training volumes with noisy “preliminary” labels, 40 high-quality training volumes, and 60 high-quality internal test volumes with all 12 organs annotated. Initial labels for all 5,731 cases were generated by a state-of-the-art automatic segmentation model plus a junior annotator, using the LIFEx tool, while the 100 high-quality cases were re-annotated and verified by two senior PET imaging experts. The dataset is thus explicitly structured around discrepant annotation quality.

The dataset ECPC-IDS addresses a narrower disease setting: endometrial cancer PET/CT for semantic segmentation and hypermetabolic-region detection (Tang et al., 2023). Its segmentation subset includes 1,193 PET images split into 476 training, 477 validation, and 240 testing. Its detection subset includes 1,093 PET images split into 965 training, 108 validation, and 120 testing. PET images were acquired on a GE Discovery PET/CT 690 with $Q_i = \frac{1}{5}\sum_{r=1}^5 q_{i,r},$3F-FDG, 3.7 MBq/kg intravenous injection, 60 minutes rest prior to scan, 3D mode, 1.5 min per bed position (7–8 beds), matrix 192 × 192, and reconstruction by OSEM with 2 iterations, 24 subsets, and a 6.4 mm Gaussian smoothing filter. After PET–CT registration and cropping, PET slices are resampled to 512 × 512; pre-cropped slices are also exported as 8-bit PNG, and detection labels use PASCAL VOC–style XML. Semantic segmentation contours were produced in 3D Slicer v4.11 by two board-certified radiologists, with a third expert resolving discrepancies by consensus.

A third line of work uses PET/CT datasets for slice classification rather than segmentation. A multi-centric lymphoma cohort contains 246 lymphoma patients from BC Cancer, Vancouver and 220 patients from Seoul St. Mary’s Hospital, with on the order of 140,000 axial slices of PET and co-registered CT (Ahamed et al., 2024). PET tracer was $Q_i = \frac{1}{5}\sum_{r=1}^5 q_{i,r},$4F-FDG, attenuation-corrected and converted to SUV$Q_i = \frac{1}{5}\sum_{r=1}^5 q_{i,r},$5; CT was diagnostic low-dose CT in Hounsfield units. Positive slices are defined as slices that contain any part of a physician-annotated 3D lymphoma tumor volume. Annotation used 4 nuclear-medicine physicians at BCCV + 1 physician at SMHS, with STAPLE fusion for 10 BCCV cases annotated by all four BCCV readers.

The most technically important observation in that study concerns experimental design: slice-level split overestimates performance because adjacent slices from the same patient may appear in both train and test, whereas patient-level split yields more realistic, lower but more trustworthy metrics (Ahamed et al., 2024). This is a general warning for PET/CT datasets that are naturally organized as contiguous volumetric series.

5. Synthetic PET and CT-to-PET translation datasets

Some PET datasets are not acquired as conventional benchmark corpora but generated or repurposed for reconstruction and translation research. In “Synthetic PET via Domain Translation of 3D MRI”, a dataset of 56 whole-body clinical PET/MRI studies with $Q_i = \frac{1}{5}\sum_{r=1}^5 q_{i,r},$6F-FDG tracer acquired on a 3.0 T time-of-flight PET/MRI scanner (GE Signa) is used to train a 3D residual U-Net to predict physiologic PET uptake from MRI (Rajagopal et al., 2022). The split consists of 40 exams for network training, 16 exams for whole-body testing, and 20 independent pelvic exams (with matched CT) for reconstruction/quantification validation. PET and post-contrast T$Q_i = \frac{1}{5}\sum_{r=1}^5 q_{i,r},$7-weighted MRI volumes were registered into MRI image space and resampled to 1 mm isotropic voxels, and PET intensities were converted to SUV using injected dose, half-life, positron fraction, elapsed time, and patient weight.

The output of that pipeline is a synthetic PET dataset in the reconstruction sense: ToF sinograms, CTAC and MRAC $Q_i = \frac{1}{5}\sum_{r=1}^5 q_{i,r},$8-maps, and lesion simulations. Synthetic lesions are inserted at four pelvic sites selected by a radiologist: acetabulum, sacrum, rectum, nodal region, using 12 mm spheres of 8 SUV. The released data format includes ToF sinograms for each exam, corresponding CTAC and MRAC $Q_i = \frac{1}{5}\sum_{r=1}^5 q_{i,r},$9-maps, and approximately 56 whole-body cases + 20 pelvis cases, with sinograms stored in vendor-standard format. The reported use case is not report generation or segmentation but development, evaluation, and validation of PET/MRI reconstruction methods.

A second translation-oriented resource is the large-scale CT–PET dataset introduced with CPDM (Nguyen et al., 2024). It contains 3,454 patients and 2,028,628 paired 2-D CT–PET image pairs, with 250–500 slices per study on average. CT was acquired on GE Discovery 710 or GE Discovery STE PET/CT scanners with 512×512 resolution, 3.75 mm or 5 mm slice thickness, 120 or 140 kVp, and HU slope = 1.0, intercept = -1024. PET used $q_{i,r}\in\{0,1,2,3,4\}$0F-FDG, 256×256 resolution, 3.27 mm slice thickness, $q_{i,r}\in\{0,1,2,3,4\}$1 min post-injection, and CT-based attenuation correction. Each DICOM slice retains patient age, sex, weight, injected activity, and acquisition parameters, with identifiers stripped.

For experiments, the authors define a subset of 598 studies and 15,000 CT–PET pairs, uniformly downsampled/cropped to 256×256, with intensity normalization to $q_{i,r}\in\{0,1,2,3,4\}$2 and a study-wise 80:10:10 split ensuring no patient overlap. The full 2 M-image dataset is privately held by the authors’ institutions, while data samples and all source code are publicly available. Relative to PET-CT-IQA-DS or PETWB-REP, this resource is optimized for CT-to-PET image synthesis, not for subjective quality labeling or report-centric multimodal learning.

6. Non-medical uses of the PET acronym

Outside medical imaging, PET may refer to the Pascal animal classes Eye-Tracking database (Gilani et al., 2016). This dataset comprises 4,135 images from the train+val subsets of PASCAL VOC 2012 for the animal classes bird, cat, cow, dog, horse and sheep, with eye movement recordings from 40 university students under both free-viewing and visual-search conditions. The acquisition apparatus was a Tobii desktop eye-tracker, 120 Hz sampling, approximately 0.4° accuracy. After discarding the first fixation per image and invalid points, the dataset contains 28,733 fixations for free-viewing and 29,901 fixations for visual search. Fixation records are stored per image and per user as tuples

$q_{i,r}\in\{0,1,2,3,4\}$3

Released files are described as including images/, fixations_free.csv, fixations_search.csv, annotations/, and meta.json. As a baseline application, max-pooling on fixation regions slightly outperformed uniform spatial pyramid pooling, with an approximate +3% absolute gain in average class accuracy over standard max pooling.

In natural language processing, PET denotes a dataset for process extraction from natural language text (Bellan et al., 2022). It contains 45 human-annotated process descriptions, 413 sentences, and annotation layers for Activity, Activity Data, Actor, Further Specification, XOR Gateway, AND Gateway, Condition Specification, together with relations such as Sequence Flow, Uses, Actor Performer, Actor Recipient, Further Specification, Same Gateway. Annotation was performed in Inception, with data released as JSON and accessible through Hugging Face. The corpus is intended for business process extraction from text, not for imaging. The coexistence of this corpus with PET imaging datasets demonstrates that acronym-only references are insufficient in technical writing.

The broad lesson across these resources is that PET datasets differ not only in scale and annotation density but also in ontology, supervision regime, unit of observation, and legal accessibility. In medical imaging, PET datasets may provide MOS labels, SUV volumes, bounding boxes, pixel-wise masks, reports, clinical metadata, or sinograms. In other domains, the same acronym may designate gaze trajectories or process graphs. Any rigorous use of “PET dataset” therefore depends on disambiguation at the level of task and data model, not merely acronym expansion.