FDTooth: Intraoral Photographs and Cone-Beam Computed Tomography Images for Fenestration and Dehiscence Detection

Abstract

FDTooth is a comprehensive dataset designed for the automated detection of fenestration and dehiscence (FD) in anterior teeth, combining intraoral photographs and corresponding cone-beam computed tomography (CBCT) images from 241 patients aged 9 to 55 years. The dataset includes 1,800 annotated bounding boxes for intraoral photographs, with cone-beam computed tomography (CBCT) serving as the imaging modality of choice for assessing bone defects. It supports interdisciplinary dental research by enabling the development of deep learning models for non-invasive FD screening, providing an alternative assessment method that reduces radiation exposure. FDTooth serves as a valuable resource for advancing dental diagnostics, training professionals, and exploring other applications in dental and maxillofacial imaging.

Background

Fenestration and dehiscence (FD) are prevalent dental conditions that adversely affect oral health by causing bone loss, exposed tooth roots, and, in severe cases, tooth loss[1-3]. Current diagnostic tools like cone-beam computed tomography (CBCT) are precise but limited by radiation exposure and resource constraints[4]. Early signs of FD are subtle and typically assessed through subjective clinical intraoral observation (e.g., noting altered gingival contours or prominent root shapes)[5, 6]. In routine practice, clinicians initially evaluate the intraoral condition visually without invasive procedures such as palpation or radiographs; however, such subjective assessments can result in inconsistent diagnoses among clinicians of varying experience. Despite the significant implications of FD in periodontology, orthodontics, and prosthodontics, there is no publicly available dataset combining intraoral photographs with cone-beam computed tomography images. Such a dataset can accelerate the development of deep learning models, providing non-invasive and efficient diagnostic alternatives to clinicians. FDTooth aims to address this gap by offering a high-quality, annotated dataset to the dental and AI research communities, facilitating interdisciplinary advancements in diagnostics and treatment planning.

Methods

Annotation Process

The labels were independently generated by two experienced dentists. Inter-examiner reliability was assessed using Cohen's kappa test, demonstrating excellent consistency (κ > 0.9). Two weeks later, the dentists repeated the annotation process for intra-examiner reliability evaluation, again achieving high agreement (κ > 0.9). Any discrepancies between annotators were resolved through consensus discussions, and unresolved cases were excluded. Finalized consensus annotations served as the gold-standard labels.

De-identification Process

The dataset was anonymized following procedures approved by the Institutional Review Board (IRB Ref No: UW 23-632). For intraoral photographs, no personally identifiable features were present, as these images only capture the intraoral region. Therefore, anonymization focused solely on renaming the files using anonymized patient IDs. For cone-beam computed tomography (DICOM) files, all identifiable metadata—including patient name, ID, and study dates—was removed using a custom script developed by one of the authors (Marawan). File names were also replaced with anonymized IDs (e.g., 001, 002, 003, …, 241) to ensure full compliance with data protection standards.

Data Description

Detailed Folder Structure:

The FDTooth dataset is organized into three main folders:

1. Patient Information:

• An Excel-compatible CSV file listing anonymized patient IDs along with demographic data, including age and gender.

2. Patient Imaging Data:

• Intraoral Photographs: High-resolution JPEG images (5760 × 3840 pixels), organized into subfolders named by anonymized patient IDs(e.g., 001, 002, 003, …, 241).
• cone-beam computed tomography (CBCT) Images: DICOM files organized into patient-specific subfolders, each labeled with a unique anonymized ID(e.g., 001_cbct, 002_cbct, 003_cbct, …, 241_cbct).

3. Annotations:

• CSV Files: Include tooth-level annotations for each of the 12 anterior teeth per patient, indicating fenestration ('F'), dehiscence ('D'), or normal ('N') conditions. Each row corresponds to a patient, clearly structured by patient ID and standardized tooth numbering (International Dental Federation numbering system).
• JSON Files: Contain bounding box annotations compatible with the MakeSense visualization tool. These files are intended for visual review and can be imported into the MakeSense platform to display the FD status (fenestration, dehiscence, or normal) of each anterior tooth.

CSV File Contents:

Each row in the CSV annotation file includes:

1. Anonymized patient ID
2. Tooth numbers (FDI World Dental Federation notation)
3. Corresponding diagnostic labels: 'F' = Fenestration; 'D' = Dehiscence; 'N' = Normal.

Basic Dataset Statistics

• Number of patients: 241
• Number of intraoral images: 241
• Number of cone-beam computed tomography (CBCT) scans: 241
• Age: Median = 28 years (range: 9–55)
• Gender distribution:
    • Females: 185 patients (76.8%)
    • Males: 56 patients (23.2%)
• Tooth-level annotation count (across 2892 anterior teeth):
    • Fenestration (F): 780 teeth
    • Dehiscence (D): 796 teeth
    • Normal (N): 824 teeth
• Patient-level distribution by FD status:
    For clarification, each patient in the dataset is associated with one intraoral photograph and one CBCT scan, covering 12 anterior teeth per patient. The FD condition (fenestration, dehiscence, or normal) is labeled for each tooth. Therefore, the number of images equals the number of patients, and per-image classification is based on whether any of the 12 teeth shows fenestration and/or dehiscence.
    • Both fenestration and dehiscence: 145 patients
    • Only fenestration: 60 patients
    • Only dehiscence: 23 patients
    • Neither condition (normal): 13 patients

Usage Notes

FDTooth provides unique opportunities for research and clinical applications, including:

1. Development of AI Models: Facilitates the creation of deep learning algorithms for automated FD detection in intraoral photographs.
2. Educational Resource: Enhances training for dental students and professionals by providing real-world imaging data.
3. Complementary Research: Enables exploration of epidemiology, diagnostic methods, and treatment outcomes for FD. Known limitations include the dataset's focus on anterior teeth and exclusion of other dental pathologies. Researchers are encouraged to cite this dataset and consider its potential for extension to other dental conditions.

Known Limitations:

• The dataset exclusively focuses on anterior teeth; posterior teeth are not included.
• Cases with severe dental pathologies or anomalies such as impacted teeth, extensive restorations, severe crowding, or active periodontal diseases are excluded.
• cone-beam computed tomography (CBCT) imaging was conducted using a single scanner model; thus, performance of developed AI models might vary when applied to images from other cone-beam computed tomography (CBCT) scanners.
• Despite rigorous annotation procedures, inherent minor subjectivity in visual-based annotations by clinicians might exist, potentially influencing annotation consistency.

Researchers are encouraged to cite this dataset and consider these limitations when interpreting results or planning extensions to other dental conditions.

Ethics

All patients included in the FDTooth dataset were teaching patients who provided informed consent prior to treatment. For minors, consent was obtained from their parents or legal guardians. The consent stated that patient data could be used for research purposes, in compliance with ethical guidelines and anonymization procedures. All data were anonymized to ensure confidentiality by removing personal identifiers such as names, contact details, and hospital IDs. The dataset includes only anonymized patient IDs (e.g., 001, 002, 003), age, and gender, with CBCT images limited to the maxillary and mandibular regions, excluding areas above the eyes, in full compliance with ethical and data protection standards.

As this is a retrospective study using data collected between January 2010 and January 2023, additional informed consent was not required. The study was approved by the Institutional Review Board (IRB) of the University of Hong Kong/Hospital Authority Hong Kong West Cluster (IRB Reference No: UW 23-632). This study adheres to the principles of the Declaration of Helsinki, ensuring participant privacy, safety, and ethical integrity.

Conflicts of Interest

The authors declare no conflicts of interest.

References

1. Löst C. Depth of alveolar bone dehiscences in relation to gingival recessions[J]. J Clin Periodontol, 1984,11(9):583-589. doi:10.1111/j.1600-051x.1984.tb00911.x.
2. Wennström JL, Stokland BL, Nyman S, et al. Periodontal tissue response to orthodontic movement of teeth with infrabony pockets[J]. Am J Orthod Dentofacial Orthop, 1993,103(4):313-319. doi:10.1016/0889-5406(93)70011-C.
3. Sun L, Yuan L, Wang B, et al. Changes of alveolar bone dehiscence and fenestration after augmented corticotomy-assisted orthodontic treatment: a CBCT evaluation[J]. Prog Orthod, 2019,20(1):7. doi:10.1186/s40510-019-0259-z.
4. Scarfe WC, Azevedo B, Toghyani S, et al. Cone Beam Computed Tomographic imaging in orthodontics[J]. Aust Dent J, 2017,62 Suppl 1:33-50. doi:10.1111/adj.12479.
5. Shah A, Shah P, Goje SK, et al. Gingival Recession in Orthodontics: A Review[J]. Advanced Journal of Graduate Research, 2017,1(1):14-23. doi:10.21467/ajgr.1.1.14-23.
6. Larato DC. Alveolar plate fenestrations and dehiscences of human skull[J]. Oral Surgery Oral Medicine Oral Pathology, 1970,29(6):816-819. doi:10.1016/0030-4220(70)90429-9.