AIRC-LABDEN: A multi-modal dataset for dental plaque diagnosis of patients with fixed labial orthodontic appliances

Abstract

One of the critical aspects of oral hygiene management in orthodontic patients with fixed appliances is the identification of dental plaque. In this paper, we present AIRC-LABDEN, a new multi-modal dental dataset specifically designed for the detection and quantification of plaque accumulation on the teeth of patients wearing orthodontic brackets using deep learning. The dataset integrates 10,450 high-resolution intraoral images from 148 patients with their corresponding anonymized clinical records, providing demographic and treatment-related data to support comprehensive analysis. All images were acquired using professional photographic equipment and standardized imaging protocols. Each image is annotated with labels indicating the presence of plaque on individual teeth, a process involving two dentists with a minimum of five years of clinical experience. By facilitating the development of multi-modal plaque detection systems, this asset helps orthodontic patients improve their oral care routines while providing a practical foundation for embedding authentic data into dental studies and academic frameworks.

To provide a comprehensive overview of the shared data resources, the following Specifications Table details the specialized scope, file formats, and the end-to-end acquisition and processing workflow of AIRC-LABDEN. This dataset, comprising authentic clinical records from 148 patients across major medical institutions in Hanoi, Vietnam, was compiled under rigorous quality control and stringent ethical research standards.

Specifications Table

Subject	Dentistry; Computer Science
Specific subject area	Orthodontics; Dental Plaque Diagnosis; Medical Imaging; Deep Learning; Artificial Intelligence in Dentistry
Type of data	Image (Intraoral photographs); Text (Clinical records); Figure (Annotations/Labels); JSON (Metadata)
How the data were acquired	-Imaging: High-resolution intraoral images were captured using a Canon RP mirrorless camera equipped with a Canon RF 100 mm f/2.8 L Macro lens, Godox MF12 macro flash, and Dantech KD1 diffuser kit. -Clinical Data: Anonymized records were extracted from patient medical history. -Annotations: Semi-automated labeling using YOLOv8 followed by manual verification by dentists with >5 years of experience.
Data format	Raw (original images); Augmented (processed images), Analyzed (annotations in .txt format), Filtered (clinical records in .docx).
Parameters for data collection	Data collected from 148 patients with fixed orthodontic appliances, totaling 10,450 images. Images are captured from 9 distinct angles and 3 noise images per session. Strict inclusion criteria: Resolution ≥ 12 MP (6240 × 4160), ROI ≥ 512px, Sharpness ≥ 90.
Description of data collection	Intraoral photographs were collected from patients undergoing fixed orthodontic treatment. The dataset includes paired images (pre- and post-disclosing agent application), anonymized clinical records (demographics, SBI index, oral habits), and tooth-surface level plaque annotations.
Data source location	Institutions: Faculty of Dentistry (VNU University of Medicine and Pharmacy) and 108 Military Central Hospital. City/Country: Hanoi, Vietnam.
Data accessibility	Repository name: Mendeley Data Data identification number (Part 1): https://dx.doi.org/10.17632/g8yhdvgjy2.3 Data identification number (Part 2): https://dx.doi.org/10.17632/xjs4bfgzj5.3
Related research article	Not applicable (dataset descriptor)

1. Value of the Data

• •AIRC-LABDEN helps bridge the gap in orthodontic AI research. This study introduces the pioneering open-access dataset tailored for identifying dental plaque in individuals undergoing fixed labial orthodontic treatment. Unlike existing datasets that focus on clean teeth, this dataset addresses the critical research gap caused by the visual complexity of brackets and archwires, which significantly alter plaque accumulation patterns.
• •AIRC-LABDEN has large-scale and robust generalization. It is large in size, comprising 10,450 high-resolution images from 148 patients, which allows for the training of robust deep learning models. We incorporated noise into the dataset to ensure diversity. This real-world complexity prevents models from overfitting to simple tooth surfaces, thereby enabling better generalization and higher accuracy when applied to actual orthodontic clinical environments.
• •AIRC-LABDEN provides multi-modal data for enhanced predictive standards consisting of two distinct data types, namely the intraoral photography and anonymized clinical records. The integration of clinical records allows researchers to correlate visual findings with patient history. This additional layer of metadata helps raise the standard of model prediction, moving beyond simple image segmentation to more holistic, patient-specific diagnostic assessments.
• •AIRC-LABDEN ensures data being annotated by dental professionals specializing in orthodontics and cross-checked by experts with over five years of experience in the field. The dataset consists of high-resolution images.
• •AIRC-LABDEN has comprehensive demographic coverage collected from diverse subjects at major medical institutions in Vietnam, ensuring a representative sample of the Asian population. It encompasses a wide range of demographic characteristics, including gender, age groups, and varying durations of wearing braces. This diversity ensures the dataset is applicable to a broad spectrum of patients.
• •AIRC-LABDEN supports educational and clinical utility. Beyond AI development, this dataset serves as a valuable resource for dental education and clinical training. It aids clinicians and students in recognizing plaque patterns in complex orthodontic cases, supporting decision-making and improving oral hygiene management for patients with braces.
• •In orthodontic patients, relying solely on visual data is often insufficient for the accurate interpretation and diagnosis of dental plaque due to challenges such as glare from orthodontic brackets, partially occluded tooth surfaces, and indistinct plaque deposits. The integration of anonymized clinical records, including treatment duration, gingival bleeding indices, and oral hygiene practices, serves as an essential supplementary metadata layer to contextualize the patient's medical history.

2. Background

In recent years, deep learning has been increasingly applied to the detection and quantification of dental plaque. Several datasets have been developed to support this research direction, yet significant limitations remain regarding their applicability to orthodontic patients. For instance, Chen et al [1] introduced the DeepPlaq dataset, which includes over 1400 stained tooth images and has been widely used for automated plaque index prediction. Similarly, You et al [2] presented a dataset of primary-teeth intraoral photographs, demonstrating that convolutional neural networks can effectively recognize plaque patterns in pediatric dentition. Addressing the need for accessible diagnostics, Nantakeeratipat et al [3] developed a smartphone-based plaque dataset consisting of paired stained and unstained anterior-tooth photographs captured using standard mobile devices.

Beyond plaque-specific datasets, other intraoral image collections have been utilized in oral hygiene-related studies. Alfonso et al. [4] employed a plaque-stained RGB dataset comprising only 531 images to classify plaque maturity stages using YOLO-based detectors. Furthermore, while primarily focusing on gingivitis rather than plaque, Duy et al [5] contributed a gingivitis dataset consisting of 1096 high-quality images. This dataset is frequently cited in periodontal AI research, contributing to broader efforts in oral pathology detection.

Despite the availability of these datasets and numerous studies that have applied deep learning to plaque detection, none of the existing datasets include images of patients wearing fixed orthodontic appliances (braces). Brackets and archwires greatly alter the visual characteristics of tooth surfaces and influence plaque accumulation patterns. Consequently, models trained on 'clean-teeth' datasets provided by Chen et al. [1], You et al. [2], or Alfonso et al. [4] failed to ensure reliable generalization when applied to orthodontic cases. Furthermore, the prior datasets are often constrained by modest sample sizes (typically ranging from a few hundred to approximately 1400 images) and sparse annotations. These are frequently limited to simple classification or bounding boxes, lacking the detailed clinical metadata essential for comprehensive analysis.

To address these limitations, this paper introduces a large-scale multimodal dataset that significantly surpasses existing resources in both quantity and depth. Unlike the previous small-scale collections, our dataset comprises 10,450 high-resolution images obtained from 148 patients, captured from nine distinct angles per session to ensure comprehensive coverage of the dental arch. Furthermore, we incorporated three types of noise images per patient to enhance diversity, ensuring better generalization and robustness in real-world orthodontic clinical settings. To facilitate more comprehensive diagnostic assessments, the dataset includes clinical records for each patient, thereby enhancing the model's predictive standards. While prior datasets primarily relied on basic image-level labels, our dataset provides detailed annotations at the tooth-surface level. Specifically, we collected and processed labels where each tooth is geometrically segmented into four specific regions, following the methodology of Beberhold et al [6], to facilitate precise and localized plaque detection.

The dataset focuses on patients undergoing orthodontic treatment with fixed appliances, a population known to present a high burden of dental plaque accumulation during treatment. Previous studies have shown that plaque coverage in these patients can reach a mean of 41.9% ± 18.8% of tooth surfaces, with values up to 85.3% and more than one third of patients exhibiting plaque coverage exceeding 50%. More recent evidence indicates that fixed appliances are associated with an increase of approximately 10% in plaque index over time, with 40% of patients presenting elevated levels of Streptococcus mutans and Lactobacilli. These data highlight the substantial and clinically relevant plaque burden around brackets and along the gingival margin, which is difficult to quantify using conventional clinical indices. Therefore, this dataset provides valuable, representative data for the development and validation of objective, image-based and artificial intelligence–assisted methods for plaque assessment in orthodontic patients [7,8].

The dataset was curated through a rigorous semi-automated pipeline, combining AI-assisted proposals (utilizing YOLOv8) with expert validation. The annotation protocol involved the use of disclosing agents for plaque visualization, supplemented by cross-verification by two dentists with over five years of experience to ensure clinical accuracy. Notably, this is the first dataset to integrate anonymized clinical records alongside imaging data, enabling multimodal research to establish correlations between visual plaque evidence and patient history.

3. Data Description

3.1. Description of dataset folders

The dataset is organized into four primary components: image data, clinical records, metadata, and legal documentation (Fig. 1). It comprises intraoral images collected from a total of 148 patients undergoing fixed orthodontic treatment, resulting in 10,450 images. For each patient, an anonymized clinical record and a corresponding informed consent form are provided, ensuring completeness and traceability of clinical information while maintaining patient privacy. An overview of dataset components can be found in Table 1.

Fig. 1.

The data directory.

Table 1.

An overview of dataset components.

Files Data_splits.csv Clinical_records Metadata Data Legal documents	Meaning
	Contains important features such as patient ID, patient image filename and status which provides a suggested structure for dividing the dataset into training, validation, and test sets
	Contains anonymized clinical data stored as Microsoft Word documents (patientxxxx.docx). Each file corresponds to a specific subject, providing demographic details and clinical conditions to support multi-modal analysis.
	This directory houses technical support files necessary for interpreting the annotations. Specifically, it contains teeth_position_mapping.json, which provides the schema linking numeric tooth identifiers to their anatomical positions.
	This directory contains the core machine learning assets, divided into two parallel subdirectories: images/: Stores 10,450 high-resolution intraoral photographs, and labels/: Contains the corresponding annotation files in .txt format. Each label file shares the exact filename with its paired image. This folder contains documentation ensuring ethical compliance and data transparency, including the official approval from the Institutional Review Board (IRB), and consent-form/: A subdirectory containing the informed consent forms used in the study, provided in both Vietnamese and English

3.2. Dataset

3.2.1. Images

High-resolution dental imagery was acquired using professional-grade camera and lens systems. All images were collected in adherence to rigorous protocols established by supervising dentists and are archived in the directory Dataset/data/images. For each patient, nine intraoral images were captured to guarantee comprehensive coverage, ensuring that no dental regions were omitted or occluded.

Fig. 2.

Nine-view intraoral images of the entire dentition pre-dyeing. These views are illustrated in Fig. 2, including: A: top middle; B: top left; C: top right; D: central middle; E: central left; F: central right; G: bottom middle; H: bottom left; and I: bottom right.

To enhance data diversity and ensure model robustness when deployed in real-world orthodontic clinical settings, we acquired additional noise samples under varying conditions, including motion blur, low-light illumination, and overexposure (glare) (Fig. 3).

Fig. 3.

Three types of noise images: A: Blur, B: Dark, C: Light.

Furthermore, to increase data diversity and improve the generalization capability of AI models, a systematic data augmentation process was applied to the original images. Five specific transformation techniques were employed:

• •Brightness Down: Decreasing luminosity to simulate underexposed conditions.
• •Brightness Up: Increasing luminosity to simulate overexposed conditions.
• •Flip Horizontal: Mirroring the image to handle asymmetry.
• •Rotate Left 15°: Simulating minor head or camera tilts.
• •Rotate Right 15°: Simulating minor head or camera tilts.

3.2.2. Labels

The annotations represent diagnosed plaque accumulation zones on the teeth of orthodontic patients, determined by experienced dentists and archived in the directory Dataset/data/labels. For each tooth, the surface is partitioned into four zones [6] for plaque localization, as described in Table 2.

Table 2.

Teeth plaque levels labels.

Region	Symbol	How to determine
Gingival	G	Area measured from the tooth cervical margin (cementoenamel junction, CEJ) to the gingival edge of the bracket.
Incisal	I	Area measured from the bracket’s incisal/occlusal edge to the incisal edge of the incisors or to the external occlusal outline of the posterior teeth.
Mesial	M	Area measured from the bracket’s mesial edge to the mesial outline adjoining regions G and I..
Distal	D	Area measured from the bracket’s distal edge to the distal outline adjoining regions G and I.

To ensure accurate anatomical identification and support clinical analysis, the tooth position identifiers recorded in the annotation files must be interpreted using a standardized reference system. We adhere to the FDI World Dental Federation notation (ISO 3950) [9] for tooth numbering. Table 3 provides a comprehensive mapping schema, linking the dataset's internal numerical identifiers to their corresponding specific anatomical positions on the dental arch.

Table 3.

Teeth position mapping.

Teeth position	15	14	13	12	11	21	22	23	24	25
ID mapping	5	4	3	2	1	6	7	8	9	10
Teeth position	45	44	43	42	41	31	32	33	34	35
ID mapping	20	19	18	17	16	11	12	13	14	15

In the label file, for each tooth that is clearly visible, we draw bounding boxes around the four peri‑tooth regions and assign distinct labels indicating the presence or absence of plaque in a five-column format (the presence of plaque (0/1), X and Y coordinates of the bounding box center, width and height of the bounding box, and the tooth position ID).

3.3. Clinical records

To support multi-modal analysis and provide a holistic view of the patient's oral health status, each subject in the dataset is associated with a detailed clinical record, stored as an anonymized Microsoft Word document (patientxxxx.docx) in the dataset/clinical_records folder. These records contain three primary categories of information:

• •Administrative and Demographic Data: This section includes the patient’s age, gender, and the specific duration of fixed orthodontic treatment (e.g., <3 months, 6–12 months).
• •Periodontal Health Assessment (Sulcus Bleeding Index - SBI): Recognizing that dental plaque is the primary etiological factor for gingival inflammation, we incorporated the Sulcus Bleeding Index (SBI) into the dataset. For each patient, the bleeding response on probing is evaluated on a scale from 0 (healthy) to 5 (severe inflammation with spontaneous bleeding) across specific teeth. These scores provide critical clinical context, as the severity of gingival inflammation often correlates directly with the history and intensity of plaque accumulation.
• •
  Oral Hygiene Habits: The records meticulously document the patient's daily self-care routines, which are essential for understanding plaque control behaviors. Data fields include:

  • •
    Brushing habits: Frequency (times/day), duration per session, and toothbrush replacement interval.
  • •
    Equipment: Type of toothbrush used (manual, electric, ultrasonic, or orthodontic-specific) and type of toothpaste (fluoride/non-fluoride).
  • •
    Adjunctive Aids: The usage of supplementary cleaning tools such as dental floss, interdental brushes, oral irrigators (water flossers), and mouthwash.

The inclusion of anonymized clinical records substantially enhances the value of this dataset beyond image-only analysis. In fixed orthodontic treatment, plaque assessment is not determined solely by image appearance, because plaque accumulation reflects not only visible surface findings but also gingival inflammatory status, duration of exposure to plaque-retentive appliances, and oral hygiene behavior. Consequently, image data alone may be insufficient in challenging situations such as bracket-related glare, partial obscuration of cervical or interproximal surfaces, subtle plaque deposits in unstained images, and variability in image quality. The structured variables strengthen the dataset for multimodal AI development by improving interpretation in ambiguous cases, supporting risk stratification, and enhancing the clinical relevance of model outputs. Accordingly, the dataset is valuable not only for image-based classification or segmentation, but also for the development of context-aware and clinically applicable AI systems in orthodontic care.

3.4. Legal documents

To ensure transparency and substantiate adherence to ethical standards during clinical data acquisition, the Dataset/legal documents directory archives digital copies of the regulatory approvals and informed consent frameworks employed in this study.

3.5. Metadata

The metadata directory facilitates the programmatic interpretation of the dataset's annotation files. It contains a single JSON file, Dataset/metadata/teeth_position_mapping.json, which serves as a machine-readable dictionary for the tooth indexing system.

3.6. Recommended dataset split

To ensure flexibility for various training methodologies and strictly prevent patient-level data leakage, all images are stored in a unified directory accompanied by the data_splits.csv metadata file. This file not only accurately maps each image to its anonymized patient ID, but also provides a reference Train/Validation/Test split ratio (Fig. 4).

Fig. 4.

Recommended Split.

3.7. Data distribution

For the assessment of oral hygiene status, the dentition was divided into six sextants following the classification proposed by Beberhold et al [6]. Each sextant represents a defined anatomical region of the maxillary or mandibular arch as defined in Table 2. This division enables a systematic evaluation of plaque distribution across both arches and facilitates region-specific comparisons. These visibility-based tooth regions and their corresponding imaging views are summarized in Table 4, and the six regions were therefore organized in accordance with both practical image acquisition conditions and the objectives of this study.

Table 4.

Dental arch segmentation by sextants.

Sextant	Description
S1 (First)	Maxillary molar and premolar region; first quadrant (R13–15)*
S2 (Second)	Maxillary canine and incisor region; first/second quadrants (R12–22)*
S3 (Third)	Maxillary premolar and molar region; second quadrant (R23–25)*
S4 (Fourth)	Mandibular molar and premolar region; third quadrant (R33–35)*
S5 (Fifth)	Mandibular canine and incisor region; third/fourth quadrants (R32–R42)*
S6 (Sixth)	Mandibular premolar and molar region; fourth quadrant (R43–R45)*

^⁎FDI tooth-numbering system [9].

Plaque accumulation on the labial surfaces of teeth with multibracket appliances was quantified using the OPI. Each tooth was visually assessed, and the extent of plaque coverage was scored on a five-point scale (0–4) according to Beberhold et al [6]:

• •Score 0: The bracket base circumference is entirely free of plaque.
• •Score 1: Plaque accumulation is isolated to a single surface adjacent to the bracket.
• •Score 2: Plaque is detected on two distinct surfaces at the bracket periphery.
• •Score 3: Biofilm presence is observed on three tooth surfaces surrounding the bracket.
• •Score 4: Accumulation covers all four surfaces around the bracket base, and/or is accompanied by clinical signs of gingival inflammation.

4. Research design

4.1. Experimental workflow diagram

The systematic workflow for data acquisition, labeling, and image captioning preparation is illustrated in Fig. 5. To ensure high-quality and uniform data, all experimental steps followed rigorous protocols concerning camera orientation, illumination settings, and patient alignment.

Fig. 5.

Experimental design flowchart.

This workflow outlines the end-to-end process of collecting, labeling, and validating intraoral images for research. The dentist explains the study to the patient and requests consent for intraoral photography. If the patient agrees, informed consent is obtained and a clinical examination is performed. Intraoral images are captured from nine angles before and after applying a plaque-disclosing agent. A clinical study record and image sets are then uploaded to the labeling and data management system, where images and metadata are stored. Automated quality checks are applied to ensure images meet predefined criteria. Dentists label plaque regions on the post-disclosing images, and the labeling results are reviewed by a senior dentist with at least five years of clinical experience. After successful review, the validated labelled dataset is exported for research use.

4.2. Data acquisition process

The dataset comprises intraoral imagery acquired from patients at the Faculty of Dentistry (University of Medicine and Pharmacy) and Central Military Hospital 108 during the period from October to November 2025 (Fig. 6). Informed consent was obtained from all participants following a detailed explanation of the study's goals and procedures. Prior to imaging, teeth were thoroughly dried and stained with disclosing dye. To ensure optimal visibility of the tooth crowns and marginal gingiva, a mouth gag was utilized. High-resolution captures were performed using a Canon RP mirrorless camera equipped with a Canon RF 100 mm f/2.8 L Macro lens, supported by a Godox MF12 macro flash and a Dantech KD1 diffuser kit for uniform illumination. Each patient will undergo 2 imaging sessions, before applying dye and after applying dye. 9 images of 9 different angles will be captured for each session (Table 5).

Fig. 6.

Capturing image procedure in the clinic.

Table 5.

Description for setting up angle.

No.	Evaluation Criteria	Requirement Description
I	Frontal Intraoral View
1	Patient posture	Bite in maximum intercuspation with direct incisal edge–to–incisal edge contact; maintain the position for ≥ 2 s; no head rotation or tilt > 5°
2	Instruments	Use a suitable cheek retractor to ensure adequate exposure of the required tooth regions within the image frame.
3	Camera angle	Camera pitched 0° ± 5° relative to the occlusal plane
4	Image coverage	Fully capture the four maxillary incisors (FDI: 11, 12, 21, 22) and the four mandibular incisors (FDI: 32, 31, 41, 42). The entire clinical crown from incisal edge to cervical margin must be visible.
II	Frontal Upper Intraoral Oblique View
5	Patient posture	Bite in maximum intercuspation with direct incisal edge–to–incisal edge contact; maintain the position for ≥ 2 s; no head rotation or tilt > 5°
6	Instruments	Use a suitable cheek retractor to ensure adequate exposure of the required tooth regions within the image frame.
7	Camera angle	Camera pitched upward 35° ± 5° relative to the occlusal plane
8	Image coverage	Fully capture the four maxillary incisors (FDI: 11, 12, 21, 22) and the four mandibular incisors (FDI: 32, 31, 41, 42). The entire clinical crown from incisal edge to cervical margin must be visible.
III	Frontal Lower Oblique Intraoral View
9	Patient posture	Bite in maximum intercuspation with direct incisal edge–to–incisal edge contact; maintain the position for ≥ 2 s; no head rotation or tilt > 5°
10	Instruments	Use a suitable cheek retractor to ensure adequate exposure of the required tooth regions within the image frame.
11	Camera angle	Camera pitched downward 35° ± 5° relative to the occlusal plane
12	Image coverage	Fully capture the four maxillary incisors (FDI: 11, 12, 21, 22) and the four mandibular incisors (FDI: 32, 31, 41, 42). The entire clinical crown from incisal edge to cervical margin must be visible.
IV	Left Intraoral Views
13	Patient posture	Bite in maximum intercuspation with direct incisal edge–to–incisal edge contact; head turned 30° to the right.
14	Instruments	Use a suitable cheek retractor to ensure adequate exposure of the required tooth regions within the image frame.
15	Image coverage	Maxillary left canine and premolars (FDI: 23, 24, 25) and mandibular left canine and premolars (FDI: 33, 34, 35) visible from incisal edge to cervical margin.
16	Camera axis	Camera pitched 0° ± 5° relative to the occlusal plane
V	Left Upper Oblique Intraoral Views
17	Patient posture	Bite in maximum intercuspation with direct incisal edge–to–incisal edge contact; head turned 30° to the right.
18	Instruments	Use a suitable cheek retractor to ensure adequate exposure of the required tooth regions within the image frame.
19	Image coverage	Maxillary left canine and premolars (FDI: 23, 24, 25) and mandibular left canine and premolars (FDI: 33, 34, 35) visible from incisal edge to cervical margin.
20	Camera axis	Camera pitched upward 35° ± 5° relative to the occlusal plane
VI	Left Lower Oblique Intraoral Views
21	Patient posture	Bite in maximum intercuspation with direct incisal edge–to–incisal edge contact; head turned 30° to the right.
22	Instruments	Use a suitable cheek retractor to ensure adequate exposure of the required tooth regions within the image frame.
23	Image coverage	Maxillary left canine and premolars (FDI: 23, 24, 25) and mandibular left canine and premolars (FDI: 33, 34, 35) visible from incisal edge to cervical margin.
24	Camera axis	Camera pitched downward 35° ± 5° relative to the occlusal plane
VII	Right Intraoral Views
25	Patient posture	Bite in maximum intercuspation with direct incisal edge–to–incisal edge contact; head turned 30° to the left.
26	Instruments	Use a suitable cheek retractor to ensure adequate exposure of the required tooth regions within the image frame.
27	Image coverage	Maxillary right canine and premolars (FDI: 13, 14, 15) and mandibular right canine and premolars (FDI: 43, 44, 45) visible from incisal edge to cervical margin.
28	Camera axis	Camera pitched 0° ± 5° relative to the occlusal plane
VIII	Right Upper Oblique Intraoral Views
29	Patient posture	Bite in maximum intercuspation with direct incisal edge–to–incisal edge contact; head turned 30° to the left.
30	Instruments	Use a suitable cheek retractor to ensure adequate exposure of the required tooth regions within the image frame.
31	Image coverage	Maxillary right canine and premolars (FDI: 13, 14, 15) and mandibular right canine and premolars (FDI: 43, 44, 45) visible from incisal edge to cervical margin.
32	Camera axis	Camera pitched upward 35° ± 5° relative to the occlusal plane
VIV	Right Lower Oblique Intraoral Views
33	Patient posture	Bite in maximum intercuspation with direct incisal edge–to–incisal edge contact; head turned 30° to the left.
34	Instruments	Use a suitable cheek retractor to ensure adequate exposure of the required tooth regions within the image frame.
35	Image coverage	Maxillary right canine and premolars (FDI: 13, 14, 15) and mandibular right canine and premolars (FDI: 43, 44, 45) visible from incisal edge to cervical margin.
36	Camera axis	Camera pitched downward 35° ± 5° relative to the occlusal plane

4.3. Preprocessing data

To ensure dataset quality and consistency, all images underwent a stringent preprocessing and quality-control workflow after acquisition. First, the entire dataset was standardized to a uniform orientation: vertically captured images were rotated 90° to achieve a horizontal (landscape) format. Subsequently, all images were resized to a fixed resolution of 5536 × 4160 pixels. Bicubic interpolation was used to preserve image quality and fine detail, an essential prerequisite for detecting subtle features such as dental plaque.

Following standardization, each image was analyzed using a custom-built automated assessment tool (Fig. 7). While the tool evaluates nine distinct parameters to monitor technical consistency, three mandatory criteria were established as strict inclusion thresholds. Images failing any of these three core requirements were automatically excluded from the dataset:

• •Resolution: The image resolution must be greater than or equal to 12 MP (specifically 6240 × 4160 pixels) to guarantee high-definition clarity for minute feature detection.
• •Minimum Size (ROI): To ensure sufficient spatial detail for the deep learning model, each dimension of the Region of Interest must be greater than or equal to 512 pixels.
• •Sharpness: To eliminate motion blur or focus errors, image sharpness was quantitatively evaluated using the Variance of Laplacian method, requiring a minimum score of 90.

Fig. 7.

A tool that checks nine criteria using the input image.

In addition to these mandatory checks, the tool also monitored six secondary technical parameters: File Format (validating JPG, PNG, or CR3), Aspect Ratio (4:3 or 3:2), ISO Range (target 200–400), Aperture (target f/18–f/22), and White Balance (target 4500K–5000 K). These parameters served as quality indicators to guide the photography process but were treated as flexible constraints rather than grounds for automatic exclusion.

Our labeling pipeline is designed as a semi-automated process to ensure consistency, efficiency, and clinical accuracy. It comprises two main stages: (1) automated object detection using a deep-learning model and (2) region delineation via a geometric algorithm, followed by (3) expert dental verification and final labeling.

To ensure data accuracy and labeling speed, we propose training a YOLOv8 model to detect and localize two object classes:

• •Tooth: The entire visible clinical crown encompassing the four predefined regions
• •Bracket: The orthodontic appliance affixed to the tooth surface

The output of this stage is two bounding boxes for each bracketed tooth (tooth-level and bracket-level), providing precise coordinates that serve as inputs to the subsequent stage.

Leveraging the bounding box coordinates derived from the YOLO model, we employed a formula to automatically partition the tooth surface into four distinct zones. This approach guarantees consistent delineation across the entire dataset. Let the bounding box of a tooth be defined as follows:

$$\left(x_{t,min},y_{t,min}\right),\left(x_{t,max},y_{t,max}\right)$$

and the bracket bounding box by:

$$\left(x_{b,min},y_{b,min}\right),\left(x_{b,max},y_{b,max}\right)$$

The four regions are then defined by the following formulas:

Gingival region ($B_G$): The area gingival to the bracket, bounded vertically by the tooth’s superior border and the gingival edge of the bracket, and horizontally by the bracket’s width.

$$B_G=\left(x_{t,min},y_{t,min},x_{t,max},y_{b,min}\right)$$

Incisal region ($B_I$): The area incisal/occlusal to the bracket, bounded vertically by the bracket’s incisal/occlusal edge and the tooth’s inferior border, and horizontally by the bracket’s width.

$$B_I=\left(x_{t,min},y_{b,max},x_{t,max},y_{t,man}\right)$$

Mesial region ($B_M$): The area on the mesial side of the bracket, bounded horizontally by the mesial border of the tooth and the mesial edge of the bracket, and vertically by the bracket’s height.

$$B_M=\left(x_{t,min},y_{b,mix},x_{b,mix},y_{b,max}\right)$$

Distal region ($B_D$): The area on the distal side of the bracket, bounded horizontally by the distal border of the tooth and the distal edge of the bracket, and vertically by the bracket’s height.

$$B_D=\left(x_{b,max},y_{b,min},x_{t,max},y_{b,max}\right)$$

4.4. Expert validation and final labeling

To ensure the accuracy and consistency of the annotations, we developed a custom web-based platform designed specifically for patient data management and efficient image labeling. This platform provided a unified interface for the experts to review the automated proposals generated in the previous steps and assign final diagnostic labels (Fig. 8).

Fig. 8.

The web-based annotation interface.

The ``gold standard'' for identifying dental plaque in this study was established using plaque-disclosing agents (GC Tri Plaque ID Gel) (Fig. 9). Although the objective of the dataset is to enable detection on unstained teeth, the annotation process utilized the corresponding stained images, captured immediately after the dyeing procedure, as the ground truth reference. On the web platform, annotators were able to cross-reference the unstained images with their stained counterparts to precisely locate and label plaque accumulation that might be subtle or invisible to the naked eye.

Fig. 9.

The 9 post-staining views correspond directly to the pre-staining angles.

The validation and labeling process was conducted by two independent dentists, each possessing a minimum of five years of clinical experience in orthodontics. The workflow adhered to a rigorous protocol:

• •Geometric Verification: Experts reviewed the bounding boxes for the four tooth regions (G, I, M, D) generated by the geometric algorithm. Any misalignments caused by irregular tooth positioning were manually corrected via the web interface.
• •Plaque Labeling: Using the stained images as the visual standard, experts assigned binary labels (1: Plaque, 0: No Plaque) to each region.
• •Cross-Validation: To guarantee data reliability, a cross-check mechanism was implemented. A subset of the data underwent a double-blind peer review, any discrepancies in labeling were flagged and resolved through consensus to ensure high inter-rater reliability for the final dataset.

Limitations

The use of multiple imaging angles and standardized acquisition procedures enhances representativeness when using this dataset for research related to multi-modal deep learning for Dental Plaque Diagnosis of Patients with Fixed Labial Orthodontic Appliances. However, the dataset is geographically limited (Vietnam), which may affect generalizability across populations. Therefore, further studies should enhance the generalization by more case studies.

Ethics Statement

All data collection procedures were conducted under the approval of the 108 Military Central Hospital, Hanoi, Vietnam (7092/GCN-BV). Patients were informed about the study’s objectives and data collection process and provided written consent to participate.

CRediT Author Statement

Chi Nguyen Anh: Conceptualization, Research Survey, data collection, Project Management, Data Labeling, Cross Annotation Validating, Writing Initialdraft.; Hai Pham Nhu: Data Collection, Resources Allocation, Data Labeling, Cross Annotation Validating, Review and Editing draft. Lam Vu Ngoc: Project Management, Data Collection, Resources Allocation.; Hoang Duy Bui: Data Pre-processing, Data Analysis, Writing Draft.; Quan Hong Phung: BuildingApplication, Data Analysis, Writing Draft.; Duc Minh Nguyen: Building Application, Data Analysis.

Data Availability

• Mendeley DataA new multi-modal dataset for Dental Plaque Diagnosis of Patients With Fixed Labial Orthodontic Appliances (Part 1 of 2) (Original data).

• Mendeley DataA new multi-modal dataset for Dental Plaque Diagnosis of Patients With Fixed Labial Orthodontic Appliances (Part 2 of 2) (Original data).