Deep learning-based dental plaque detection on primary teeth: a comparison with clinical assessments

Wenzhe You, Aimin Hao, Shuai Li, Yong Wang, Bin XiaResearch 2020 · DOI

Introduction

Dental plaque is the primary etiological factor for common oral diseases, including dental caries, gingivitis, and periodontitis. For children, the detection and control of plaque are vital for long-term oral health. However, identifying plaque can be challenging for parents and even clinicians without the use of disclosing agents—dyes that stain plaque to make it visible. While effective, these agents are often messy, have an unpleasant taste, and can be off-putting for young patients.

Recent advancements in artificial intelligence (AI) and computer vision offer a potential solution. This study, conducted by researchers from Peking University and Beihang University, evaluates a deep learning-based model designed to detect dental plaque on primary teeth using standard digital photography, aiming to provide a non-invasive and objective diagnostic tool.

Methodology

The researchers developed an AI model based on the DeepLabV3+ architecture, a state-of-the-art convolutional neural network (CNN) framework designed for semantic segmentation. Semantic segmentation allows the model to classify each pixel in an image, effectively "mapping" the boundaries of dental plaque.

Training and Ground Truth

The model was trained using a dataset of 886 photos of primary teeth from 86 children (ages 5–8). To establish a reliable "ground truth" for the AI to learn from, the researchers employed a two-step process:

  1. High-resolution photos were taken of the teeth in their natural state.
  2. A disclosing agent was applied, and a second set of photos was taken to reveal the exact location of the plaque.
  3. Using LabelMe software, the stained areas from the second set were mapped onto the original "clean" photos.

This allowed the AI to learn the subtle visual characteristics of plaque—such as texture and color variations—that are often invisible to the untrained eye but correspond to the areas highlighted by the dye.

Evaluation Metrics

The primary metric used to evaluate performance was the Mean Intersection-over-Union (MIoU). This metric calculates the overlap between the AI’s predicted plaque area and the actual ground truth area. An MIoU of 1.0 would indicate a perfect match.

Comparative Analysis

To validate the clinical feasibility of the model, the researchers compared the AI’s performance against an experienced pediatric dentist with 20 years of clinical practice. The comparison was conducted across two distinct validation sets:

  1. High-Resolution Set: 98 photos taken with a digital camera (Canon EOS 60D).
  2. Intraoral Set: 102 photos taken with a standard intraoral camera (TPC Ligang), which typically produces lower-resolution images.

Furthermore, to test human subjectivity, the dentist was asked to re-evaluate the high-resolution set after a one-week interval to measure intra-examiner consistency.

Results

The AI model demonstrated robust performance across all tests, achieving results comparable to, and in some cases exceeding, the human expert.

  • General Performance: On the initial test set, the AI achieved an MIoU of 0.726 ± 0.165.
  • High-Resolution Comparison: The AI achieved an MIoU of 0.736, while the pediatric dentist achieved 0.695.
  • Intraoral Comparison: On lower-resolution images, the AI maintained an MIoU of 0.724, whereas the dentist’s accuracy dropped to 0.652.
  • Consistency: While the dentist’s assessment varied slightly after one week (dropping from 0.695 to 0.689), the AI provided perfectly reproducible results.

Statistical analysis using paired t-tests confirmed that there was no significant difference (P > .05) between the AI model and the human specialist, indicating that the AI has reached a "clinically acceptable" level of diagnostic accuracy.

Discussion and Future Directions

The study highlights several key advantages of AI-based plaque detection. First, the model is highly consistent, eliminating the subjectivity and fatigue that can affect human clinicians. Second, the AI’s resilience to lower-resolution images suggests it could be effective using standard consumer hardware.

Limitations

Despite these promising results, the researchers noted several limitations:

  • Data Volume: While 886 images provided a strong foundation, larger datasets are needed to improve the model's ability to generalize across diverse dental conditions.
  • Explainability: As with many deep learning models, the "black box" nature of the CNN makes it difficult to determine exactly which visual features the AI uses to identify plaque.
  • Scope: The current study focused primarily on the labial (front) surfaces of the teeth.

Conclusion

The ultimate goal of this research is to integrate the AI model into a mobile application. Such a tool would allow parents to monitor their children's oral hygiene at home using a smartphone or a low-cost home intraoral camera. By providing real-time, objective feedback on brushing efficacy, this technology could significantly improve pediatric oral health outcomes and reduce the prevalence of dental caries.

Summary updated

Abstract

Abstract Background: Dental plaque is the cause of many common oral diseases (e.g., caries, gingivitis, and periodontitis). Therefore, the detection and control of dental plaque are of great significance to children’s oral health. The objectives of this study are to design an artificial intelligence (AI) model based on deep learning to detect dental plaque on primary teeth and evaluate the diagnostic accuracy of the AI model. Methods : A conventional neural network (CNN) framework was adopted, and a total of 886 photos of primary teeth taken by an intraoral camera (1280×960 pixels; TPC Ligang, Shenzhen, China) were used to train the AI model. To validate the clinical feasibility, 98 photos of primary teeth taken by the intraoral camera were assessed by the AI model. Additionally, tooth photos were acquired using a digital camera (3216×2136 pixels, Canon EOS 60D, Japan). One experienced pediatric dentist looked at these photos and marked the regions the photos containing dental plaque. Then, a plaque-disclosing agent was applied, and the areas of dental plaque were identified. After a week, the dentist drew the dental plaque area on the 98 photos taken by the digital camera a second time to evaluate the consistency of manual diagnosis. In another round of comparison, 102 photos of primary teeth taken by the same intraoral camera were marked to denote the dental plaque area obtained both by the AI model and by the pediatric dentist to evaluate the diagnostic ability of each approach based on photos with lower-resolution (fewer pixels) than those taken by the digital camera. The mean intersection-over-union (MIoU) metric was employed to indicate the detection accuracy. Results : The MIoU for the detection of dental plaque on the tested tooth photos was 0.726 ± 0.165.The MIoU of the dentist when diagnosing the 98 photos taken by the digital camera for the first time was 0.695 ± 0.269, while after one week, the MIoU of the dentist was 0.689 ± 0.253. Compared to the dentist, the AI model demonstrated a higher MIoU (0.736 ± 0.174), and the results did not change after a week. When both the dentist and the AI model assessed the 102 photos taken by the intraoral camera. the MIoU of the pediatric dentist was 0.652 ± 0.195, and the MIoU of the AI model was 0.724 ±0.159. The results of a paired t-test found no significant difference between the AI model and human specialist (P > .05) in diagnosing dental plaque on primary teeth. Conclusions : The AI model showed clinically acceptable performance in detecting dental plaque on primary teeth compared with an experienced pediatric dentist. This finding illustrates the potential application in adopting such AI technology for helping children improve their oral health.
BibTeX
@inproceedings{you2020deep,
  author = {Wenzhe You and Aimin Hao and Shuai Li and Yong Wang and Bin Xia},
  title = {{Deep learning-based dental plaque detection on primary teeth: a comparison with clinical assessments}},
  booktitle = {Research},
  year = {2020},
  doi = {10.21203/rs.2.21027/v2},
}
Figure