Desperado1103/Pandent_Sample

PanDent Sample Release

This repository provides a 300-case sample of PanDent, including 240 cases from the public-source portion of the dataset and 60 de-identified in-house cases.

The sample is released to facilitate inspection of image quality, structured clinical annotations, annotation format, and structure--language correspondence.

The complete PanDent dataset contains 9,524 dental panoramic radiographs, including 9,019 public-source cases and 505 in-house cases.

Dataset Composition

The eligible pools used for the current sample contain:

• —609 public-source cases
• —478 in-house cases

From these pools, 240 public-source cases and 60 in-house cases are selected using a fixed and reproducible sampling procedure.

The complete PanDent annotations and all images eligible for redistribution are planned for release upon publication, subject to the permissions and licensing terms of the original data sources.

Clinical Annotation Schema

PanDent provides 21 structured clinical attributes covering tooth-level findings and regional radiographic abnormalities.

Detailed definitions of the attributes and their canonical representation are provided in:

Annotation Schema

Clinically meaningful localization and descriptive information associated with these attributes is retained. Additional free-text clinical observations, when available, are stored separately from the structured attributes.

Repository Structure

The main files are:

• —records.jsonl: joined records for Hugging Face loading
• —annotations.jsonl: canonical structured annotations
• —reference_reports.jsonl: corresponding reference reports
• —metadata.csv: public sample metadata and image paths
• —dataset_info.json: dataset version and release information
• —images/public_source/: images from the public-source portion
• —images/in_house/: de-identified in-house images
• —statistics/: sampling and attribute statistics
• —documentation/annotation_schema.md: annotation definitions
• —documentation/DATA_USAGE.md: data usage and redistribution terms

No original patient identifiers or private case-linkage information are included in the public release.

Loading the Dataset

from pathlib import Path
from datasets import load_dataset, Image

root = Path("pandent_sample_release").resolve()

dataset = load_dataset(
    "json",
    data_files=str(root / "records.jsonl"),
    split="train"
)

dataset = dataset.map(
    lambda r: {
        "image": str(root / r["image_path"])
        if r["image_path"] else None
    }
)

dataset = dataset.cast_column("image", Image())

print(dataset[0]["attributes"])

All image paths are stored relative to the dataset root.

Reference Reports

Reference reports are constructed from the reviewed structured clinical annotations using the clinician-guided report-construction pipeline.

The corresponding report-generation code and template definitions are released together with the PanDent codebase. The reports included in this sample are frozen to ensure reproducibility.

Sampling

The public-source and in-house subsets are sampled independently using a fixed random seed of 42.

Sampling is designed to approximately preserve the natural distribution of the 21 clinical attributes while maintaining broad coverage of the annotated findings.

No artificial class balancing or rare-class oversampling is applied.

Additional sampling statistics are available in:

Sampling Report

Ethics and De-identification

All in-house panoramic radiographs are de-identified before release, and patient-identifying information is removed from the public dataset.

The in-house data were collected and processed under ethical approval from the Institutional Review Board of The University of Hong Kong/Hospital Authority Hong Kong West Cluster (UW 23--583).

Only allowlisted clinical annotations and newly generated public sample IDs are included in the released files.

Licensing and Redistribution

Images originating from public-source datasets remain subject to the licensing and redistribution terms of their original data sources.

In-house data, PanDent annotations, generated reports, and associated code are distributed according to the terms described in:

Data Usage

Users should review the relevant source-specific terms before redistributing individual dataset components.

Citation

If you use PanDent in your research, please cite:

@article{li2026pandent,
  title={PanDent: Toward Comprehensive Tooth-Level Structure-Language Consistency in Dental Radiology},
  author={Li, Xiaohan and Liu, Xinyu and Liu, Chang and Yeung, Sum Wing Au and Liu, Jun and Yuan, Yixuan and Chen, Hui},
  journal={arXiv preprint arXiv:2607.27378},
  year={2026}
}