PANDENT: A Tooth-Level Dental Panoramic Radiography Benchmark

PANDENT is a clinically grounded benchmark for multimodal learning in dental panoramic radiography, also known as orthopantomogram (OPG) imaging. It is designed to evaluate whether models can produce clinically meaningful reports that remain consistent with expert-validated, tooth-level findings.

The dataset contains 9,524 radiograph-report pairs, including 9,024 training cases and a held-out 500-case test benchmark.

Important: PANDENT is a research resource. It is not intended for clinical diagnosis, treatment planning, or direct patient care.

Dataset Summary

Item	Description
Imaging modality	Dental panoramic radiographs / OPG
Total cases	9,524
Training split	9,024 cases
Test split	500 cases
Annotation granularity	Tooth-level findings using FDI numbering, plus region-level findings
Clinical attributes	21 reportable finding dimensions
Reports	Clinician-guided, template-generated English radiology reports
Access	Gated dataset with manual approval

The final release was curated from 9,854 raw annotated records. Cases with invalid or unresolved image paths were removed. The resulting dataset includes 9,019 public-source cases and 505 de-identified in-house cases.

Data Curation and Annotation

PANDENT combines multiple public-source OPG batches with a de-identified in-house batch. The public-source component includes the source batches reported in the accompanying paper, including ToothFairy, DENTEX, Mandible, Humans, and NTU.

Only permanent-dentition OPG cases were retained. Mixed-dentition and non-dental images were excluded.

Structured clinical annotations were produced by four professional dentists with more than ten years of clinical experience. A randomly sampled subset was further validated by an oral and maxillofacial radiology expert. The released reports are not raw clinical reports: they are generated from expert-validated structured findings using clinician-defined reporting logic and a radiologist-authored template bank.

Annotation Schema

PANDENT represents both tooth-index-aware and region-level findings.

Tooth-index-aware findings use the two-digit FDI permanent-tooth numbering system and include:

• missing teeth;
• partially erupted or impacted teeth;
• caries;
• periapical lesions;
• furcation involvement;
• metallic and non-metal restorations;
• prior root-canal treatment;
• artificial crown and bridge prostheses;
• non-metal post-and-core structures;
• localized angular bone loss.

Region-level findings include generalized crestal bone loss in the maxilla or mandible, maxillary and mandibular structural abnormalities, maxillary sinus abnormalities, and temporomandular joint abnormalities.

Intended Uses

PANDENT may be used for:

• dental radiology report generation research;
• evaluation of structure-language consistency in multimodal models;
• tooth-level clinical finding localization and prediction;
• supervised fine-tuning and evaluation of dental vision-language models;
• non-commercial academic and educational research.

Out-of-Scope Uses

PANDENT must not be used for:

• clinical diagnosis, treatment planning, or patient-care decisions;
• attempts to identify, profile, or link any patient to external information;
• redistribution of raw images, annotations, reports, or any substantial portion of the dataset;
• commercial use without prior written permission;
• uses that conflict with applicable ethics approval, institutional policy, law, or original-source terms.

Access and Data Use Agreement

Access is granted on a case-by-case basis. By downloading or using PANDENT, users agree to:

1. use the dataset only for an approved research or educational purpose;
2. comply with the LICENSE and Data Use Agreement in this repository;
3. not redistribute the dataset or provide access to unauthorized parties;
4. not attempt re-identification;
5. not use the dataset for clinical deployment or commercial purposes without written authorization;
6. cite PANDENT and the original data sources in resulting publications.

Dataset maintainers may reject or revoke access when these conditions are not met.

Privacy and Ethics

The study was conducted under institutional review board approval or an equivalent ethical review. Patient data were anonymized and handled in accordance with institutional and ethical guidelines.

Users remain responsible for complying with their own institution's ethics, privacy, security, and data-governance requirements.

Repository Structure

Update this section after uploading the files:

PANDENT/
├── README.md
├── LICENSE
├── SOURCE_LICENSES.md
├── train/
│   ├── images/
│   └── annotations/
└── test/
    ├── images/
    └── annotations/