Description:FIELD OF INVENTION
The invention presents an automated methodology for precise dental plaque detection on primary teeth using advanced imaging technologies, such as intraoral cameras or digital radiographs, combined with AI techniques like deep learning and image processing. This system enhances early plaque detection, improves diagnostic accuracy, and offers personalized oral health monitoring, providing a valuable tool for pediatric dentistry and preventive care.
BACKGROUND OF INVENTION
Dental plaque accumulation on primary teeth is a significant factor in the development of early childhood dental issues such as cavities, gingivitis, and other oral health problems. Detecting plaque accurately and promptly is crucial for implementing preventive measures. However, traditional methods of plaque detection, such as visual inspection or manual disclosing agents, are time-consuming, subjective, and may miss subtle plaque formations, especially in young children whose cooperation can be limited during dental examinations. Additionally, these conventional methods often fail to detect plaque at an early stage, when intervention can be most effective. With advancements in dental imaging, technologies like intraoral cameras, digital radiographs, and fluorescence imaging have become more prevalent. However, these imaging modalities, while effective, still rely heavily on human interpretation, which can be prone to error or inconsistency. This creates the need for more reliable, objective, and automated methods for plaque detection. Artificial Intelligence (AI), particularly deep learning, has shown tremendous potential in various medical fields for automating diagnostic tasks. AI algorithms can process large amounts of image data and identify patterns that may be difficult for the human eye to detect. In the context of pediatric dentistry, combining advanced imaging techniques with AI has the potential to significantly improve the precision of plaque detection on primary teeth. This invention aims to develop an automated methodology that integrates these technologies to enhance early and accurate detection of dental plaque, thus enabling better preventive care and reducing the risk of long-term oral health complications in children.
The patent application number 201937001486 discloses a method for estimating at least one of shape position and orientation of a dental restoration.
The patent application number 202141053793 discloses a novel recalcination technique to obtain recalstone and recalsite dental gypsum products and thereof.
The patent application number 202241011527 discloses a autoclavable silicone sleeve for dental handpieces.
The patent application number 202221058934 discloses a system for detecting and quantifying a plaque/stenosis in a vascular ultrasound scan data.
The patent application number 202341039706 discloses a ocimum basilicum dental varnish composition, its method of preparation and oral use thereof.
SUMMARY
Dental plaque is a sticky, colorless film of bacteria that forms on teeth, leading to oral health problems like cavities, gum disease, and bad breath, especially in children. Early and precise detection of plaque on primary teeth is crucial for preventing these issues. However, traditional methods for plaque detection, such as visual inspection and manual plaque disclosing agents, are subjective, time-consuming, and may not identify plaque in its early stages, leading to missed opportunities for intervention. Moreover, these methods are not always suitable for young children, who may have difficulty cooperating during dental exams, further complicating accurate plaque detection.
Recent advancements in dental imaging, including intraoral cameras, digital radiographs, and fluorescence imaging, provide more detailed views of the oral cavity. However, these imaging technologies often require manual interpretation by the dentist, which can be prone to human error and inconsistencies, especially in detecting early-stage plaque. The integration of Artificial Intelligence (AI) techniques, particularly deep learning and image processing, offers a promising solution for automating plaque detection. AI can analyze complex image data, detect subtle patterns, and identify plaque more accurately and consistently than traditional methods. This invention seeks to develop an automated methodology for detecting dental plaque on primary teeth, utilizing advanced imaging and AI algorithms. By automating plaque detection, this system can provide faster, more reliable diagnoses, improving the effectiveness of preventive care in pediatric dentistry and reducing the risk of long-term oral health complications.
DETAILED DESCRIPTION OF INVENTION
Dental plaque, a sticky bacterial film that forms on tooth surfaces, is a primary contributor to periodontal diseases. If left unchecked, plaque can damage periodontal tissues and ultimately lead to tooth loss. Therefore, effective detection and control of dental plaque are essential in preventing periodontal disease.
Traditionally, plaque is detected by probing around the gingival margin with an explorer. However, as shown in Figures 1A–C, it is difficult to assess plaque levels with the naked eye without the use of dyes. To improve plaque visibility and encourage patient cooperation, erythrosine solution is commonly used to stain plaque areas, making it more distinct from the surrounding tooth surface. While this method is effective, it has several limitations. Plaque-disclosing dyes can temporarily stain the oral mucosa and lips, raising esthetic concerns. Furthermore, these techniques require specific disclosing agents and application methods, making home-use products difficult for patients to apply and interpret correctly. In clinical settings, manual plaque assessment is time-consuming and prone to human error, especially in busy environments.
To overcome these challenges, digital technologies, such as 3D imaging with intraoral scanners and fluorescence-based methods, have been explored for dental plaque detection. However, these solutions still face limitations in real-world applications, such as the need for specific staining solutions and specialized equipment. These issues underscore the need for a more convenient, automated method of detecting dental plaque.

Figure 1. Undyed and Dyed Images of Teeth with Varying Plaque Levels. (A–C) Undyed images for model training. (D–F) Dyed images showing plaque severity: (D) heavy, (E) moderate, (F) mild.
Materials and Methods
Study Approval and Participant Criteria
The study was approved by the Human Research Ethics Committee of Srinakharinwirot University (SWUEC-671045). Informed consent was obtained from all participants. One hundred dental students from the Faculty of Dentistry at Srinakharinwirot University, Bangkok, Thailand, were enrolled. Participants were required to have permanent upper anterior teeth (teeth 13–23) without restorations or fixed appliances.
Data Collection
Photographic images of the upper anterior permanent teeth were captured using a 12-megapixel smartphone camera (iPhone 13, Apple Inc., Cupertino, CA). Standardized photographic protocols were followed to ensure image consistency. The camera was mounted on a tripod, and LED lighting was used to provide uniform illumination and control ambient lighting variations. Two images were captured for each participant. The first image was taken before the application of erythrosine dye, which was applied for 30 seconds to highlight dental plaque accumulation. The second image was captured after dye application.
Image Processing and Data Augmentation
For model training, individual tooth images were isolated by cropping using the Procreate application (version 5.3, Salvage Interactive, Hobart, Australia). Each participant’s images were cropped into six separate images—one for each tooth. The images taken before and after plaque disclosure were paired. The dyed images served as ground truth labels for model training.
The level of dental plaque in the dyed images was categorized based on the percentage of the dyed area on the labial surface of the tooth. This analysis was performed using ImageJ software (version 1.54h, U.S. National Institutes of Health, Bethesda, MD). The images were converted to an 8-bit grayscale format, and thresholding was applied to isolate dyed areas. The percentage of dyed regions was calculated, with the following classifications: "mild" for less than 30% dyed area, "moderate" for 30%–60%, and "heavy" for over 60% dyed area (Figure 2).
To ensure a balanced dataset, data augmentation techniques such as flipping and rotation were applied to diversify the images used for model training. This helped improve the robustness of the model by increasing the variety of input data.

Figure 2. Image Processing and Model Classification Framework
(A) Workflow for preparing images for model training, including the capture of undyed images, application of a disclosing agent, cropping, and plaque analysis. Images are labeled into mild (<30%), moderate (30%–60%), and heavy (>60%) categories based on plaque coverage. These labeled images are used to train the AutoML models.
(B) Classification framework: The mild-moderate-heavy model categorizes images into three groups, while the acceptable-unacceptable model simplifies classification into two categories: acceptable (mild) and unacceptable (moderate and heavy).
Model Development and Training
The undyed tooth images were used to train the AutoML models on the Vertex AI platform (M125 release with Gemini 1.5 Pro, Google Cloud, Mountain View, CA). Two datasets were created for single-label classification. The first dataset contained three groups: mild, moderate, and heavy plaque levels, as determined by ImageJ analysis. The second dataset grouped mild plaque as "acceptable" and moderate/heavy as "unacceptable". The training pipeline was set to "us-central1 (Iowa)", with options configured for high accuracy, 200–300ms latency, and Google-managed encryption. The images were randomly split into training (80%), validation (10%), and test (10%) datasets. The validation set was used to tune the model's parameters, monitor performance, and prevent overfitting, while the test set was kept unseen during training.
Model Evaluation
After model development, the performance of the AutoML models was evaluated on the Vertex AI platform using statistical metrics including area under the precision-recall curve (AUPRC), precision, recall, and F1-score. These metrics were automatically analyzed to assess the model's effectiveness in detecting plaque levels.
Results
Two datasets were prepared for training. The first dataset (for the mild-moderate-heavy model) included 100 images each for mild, moderate, and heavy plaque categories. The second dataset (for the acceptable-unacceptable model) contained 196 acceptable and 103 unacceptable images. The distribution of images was adjusted to avoid excessive data augmentation.
Mild-Moderate-Heavy Model
The model was trained using 300 images: 240 for training, 30 for validation, and 30 for testing. At a confidence threshold of 0.5, the mild-moderate-heavy model achieved an average precision of 0.907, with the highest precision of 0.983 in the "heavy" category. The precision-recall curve and confusion matrix are shown in Figure 3A.
Metric Mild-Moderate-Heavy Model
Average Precision 0.907
Precision (Heavy) 0.983
Precision (Moderate) [Value]
Precision (Mild) [Value]
F1-Score [Value]

Figure 3. Model Performance and Misclassification Analysis
(A) Precision-recall curve (left), precision-recall by threshold (middle), and confusion matrix (right) for the mild-moderate-heavy model. True predictions (blue) and misclassifications (red) are shown.
(B) Example of misclassification: a true moderate plaque label (38.1% dyed area) misclassified as mild.
(C) Precision-recall curve (left), precision-recall by threshold (middle), and confusion matrix (right) for the acceptable-unacceptable model. True predictions (blue) and misclassifications (red) are shown.

This study demonstrates the potential of using Vertex AI AutoML to detect dental plaque levels on permanent upper anterior teeth based on photographic images. AutoML provides an accessible and automated solution for image classification, eliminating the need for extensive machine learning expertise. While previous research has explored machine learning for dental plaque detection, using deep learning models for segmentation, this study stands out as one of the first to apply an AutoML platform with smartphone cameras to capture standard photographs for creating a classification model. This approach simplifies model development and makes it easier for both clinicians and patients to assess dental plaque levels using commonly available technology.
The mild-moderate-heavy classification model showed promising overall performance, but label-specific accuracy varied between categories. The model performed well in identifying heavy plaque but faced challenges with moderate and mild cases. Misclassification typically occurred when the percentage of dyed area was near category thresholds, which made distinguishing between plaque levels more difficult. These challenges were likely due to the subtle changes and gradients in the images, which complicate classification. Similar difficulties have been noted in other classification tasks.
To address these challenges, we developed a binary "acceptable-unacceptable" model, which demonstrated improved performance, particularly in detecting unacceptable plaque levels. This model's strong performance is essential for clinical decision-making, as it directly informs preventive care. Simplifying the classification task to two categories led to better model performance, reflecting the preference for near-zero plaque accumulation.
The large dataset used for training contributed to the model's ability to generalize across diverse cases, improving its accuracy. However, the custom dataset used in this study, which included high-quality photographs from dental students, more closely reflected real-world conditions than pre-existing image databases. Despite this, the relatively small number of images limited the model's ability to generalize effectively. Expanding the dataset to include a more diverse group of individuals—considering different ages, demographics, dental conditions, and tooth surfaces—would enhance the model's performance. Incorporating data from various clinical settings would also help the model better adapt to different lighting conditions, smartphone models, and image quality variations, further increasing its applicability.
Although the current method of manually cropping individual teeth ensures high precision, it may not be practical in high-volume clinical settings. To address this, we are working on a new model capable of assessing plaque levels across multiple teeth in a single image. This innovation, combined with a user-friendly application, would enhance the practicality of this technology in both clinical and home settings, as well as support large-scale community screenings.
Despite the limitations, this study shows the feasibility of AutoML in real-world applications. Compared to manual plaque detection methods, which rely on visual inspection or disclosing dyes, AutoML offers several advantages: it is non-invasive, reduces time-consuming manual assessments, and eliminates the need for plaque-disclosing agents. Furthermore, advancements in AutoML allow model endpoints to be integrated into user-friendly applications, allowing for quick and easy evaluation. While the time needed to evaluate plaque may be similar to traditional dye-based methods, AutoML's efficiency and ease of use make it an attractive alternative. This simplicity allows researchers and clinicians without AI expertise to use the technology, bridging the gap between machine learning and everyday healthcare practices. As AI research in dentistry advances, this study not only showcases AutoML's capability in detecting dental plaque but also highlights its potential to become a key component in future oral health management.
The findings of this study emphasize the clinical potential of AutoML in transforming dental plaque detection practices. By offering a digital method for assessing plaque levels, this technology could significantly improve patient management. In clinical settings, automated assessments would enhance the efficiency of dental professionals and reduce chair time. Additionally, smartphone-compatible models would empower patients to take a more active role in managing their oral health through self-monitoring. This approach could be extended to large-scale community screenings, improving access to preventive care, especially in underserved populations.
In conclusion, this study demonstrates that Vertex AI AutoML is an effective tool for image-based dental plaque detection, offering high accuracy in distinguishing between plaque levels. For dental professionals, this model provides a streamlined workflow, reducing reliance on traditional dye-based methods. Its compatibility with smartphones opens new possibilities for remote monitoring and large-scale public health initiatives, improving oral health outcomes on both individual and community levels. To further enhance the model’s applicability, future studies should expand the dataset to include more diverse patient populations and clinical settings, ensuring its broad generalizability and real-world use in preventive dental care.

DETAILED DESCRIPTION OF DIAGRAM
Figure 1. Undyed and Dyed Images of Teeth with Varying Plaque Levels. (A–C) Undyed images for model training. (D–F) Dyed images showing plaque severity: (D) heavy, (E) moderate, (F) mild.
Figure 2. Image Processing and Model Classification Framework
Figure 3. Model Performance and Misclassification Analysis , Claims:1. Automated Methodology for Precise Dental Plaque Detection on Primary Teeth Using Advanced Imaging and AI Techniques claims thatthe study effectively leverages Vertex AI AutoML to develop an automated system for detecting dental plaque on primary teeth using photographic images.
2. The AI-driven methodology offers a non-invasive, automated alternative to traditional dental plaque detection methods that rely on manual inspection or plaque disclosing agents.
3. By utilizing smartphone cameras to capture dental images, the system provides an accessible solution for both clinicians and patients, requiring no specialized equipment.
4. The first model categorizes plaque levels into mild, moderate, and heavy classifications, achieving high precision in detecting heavy plaque, though facing challenges with moderate and mild levels.
5. The study develops a second, binary classification model (acceptable vs. unacceptable plaque) to address the misclassification issues observed in the first model, yielding improved performance and higher precision.
6. The study uses a custom dataset consisting of high-quality photographs taken from dental students, which more accurately reflects real-world dental conditions compared to existing image databases.
7. The study identifies the challenges in detecting subtle plaque transitions in images, particularly for moderate and mild categories, due to gradual changes and gradient appearances.
8. The second binary classification model achieved a precision of 96.4% and an F1 score of 93.1% at a 0.5 confidence threshold, demonstrating the model's strong potential for clinical use.
9. The model's ability to detect plaque levels accurately suggests its applicability in preventive dental care, allowing clinicians to make data-driven decisions and empowering patients to monitor their oral health.
10. The study indicates that AutoML-based models can be integrated into smartphone applications for remote monitoring and large-scale community screenings, addressing oral health disparities and improving public health outcomes.