Microbial Dental Plaque Analysis in Young Permanent Teeth Using Deep Learning

Ankara Medipol University

Status: Completed

Conditions: Deep Learning; Dental Plaque (Diagnosis)

Treatments: Diagnostic Test: The Difference Between The AI Model and Dentists Group; Diagnostic Test: Deep Learning Models

Study type: Interventional

Funder types: Other

Identifiers

NCT06603233

AnkaraMedipolU-DNT-BCT-01

Details and patient eligibility

About

Background: Dental plaque contributes to a number of common oral conditions such as caries, gingivitis, and periodontitis. As a result, detection and management of plaque is of great importance for the oral health of individuals. The primary objectives of this study were to design a deep learning model for the detection and segmentation of plaque in young permanent teeth and to evaluate the diagnostic accuracy of the model. Methods: The dataset contains 506 dental images from 31 patients aged 8 to 13 years. Six state-of-the-art models were trained and tested using this dataset. The U-Net Transformer model was compared with three dentists for clinical applicability using 35 randomly selected images from the test set.

Full description

Dental plaque is defined as a microbial community embedded in a matrix composed of polymers derived from bacteria and the content of saliva that develops on the surface of the teeth. Microbial dental plaque is adsorbed onto the tooth surface within seconds after dental cleaning and persists functionally. These molecules primarily exist in the fluid of the subgingival sulcus, along with saliva, and demonstrate settlement in this area. The primary etiological factor for gingivitis and periodontitis is bacterial plaque, which can lead to the destruction of gingival tissues and periodontal attachment. In children, if oral hygiene is not established immediately after tooth eruption, and regular brushing habits are not instilled, the bacterial biofilm layer can settle on the tooth surfaces and gingival margins associated with the oral environment, initiating gingival inflammation.

The early detection and treatment of periodontal diseases at the initial stages in children are clinically important, as these conditions can intensify and lead to adverse outcomes in later periods. Bacterial plaque is the primary etiological factor for gingival diseases in children. Identifying and distinguishing microbial dental plaque by patients can be challenging. Plaques can be detected through routine clinical practice using periodontal probes and/or plaque-disclosing solutions. Although these methods are widely employed, they may yield subjective results. However, these assessment methods can be cumbersome, time-consuming, and unsuccessful in noncooperative children. Additionally, plaque-disclosing solutions used for microbial dental plaque detection may temporarily stain the oral mucosa and lips. The literature also includes digital imaging analyses such as laser-induced autofluorescence spectroscopy and HIS color space for the detection of microbial dental plaque. However, the drawbacks, such as the high cost of equipment and technical standardization, limit their use .

For these reasons, this study aims to develop an affordable and easily accessible artificial intelligence (AI) model for the early and accurate diagnosis of microbial dental plaque in children. The aim is to prevent various periodontal problems and provide motivation for oral hygiene by evaluating the diagnostic and detection performance of this AI model.

With advancements in artificial intelligence for image processing, research on detecting, segmenting, and quantifying dental plaque in images captured by dental cameras has significantly increased. One study attempted to detect dental plaque using an Enhanced K-Means machine learning algorithm. Additionally, a Mask R-CNN-based dental health Internet of Things (IoT) platform was developed to classify seven different oral diseases, including dental plaque, with a perfect accuracy rate for plaque recognition, although not for segmentation.

While the U-Net model is widely regarded as successful and mainstream in the domain of biomedical image processing, there are no studies in the literature on the analysis of dental plaque with U-Net and its variants. Additionally, no studies have been encountered regarding the analysis of dental plaque in young permanent teeth of children. Hence, this study endeavors to train six state-of-the-art artificial intelligence models, incorporating variations of the U-Net model, for the purpose of dental plaque prediction in young permanent teeth of children. Subsequently, their performances are meticulously summarized and presented for comprehensive analysis. Finally, to validate the clinical feasibility of the best performing model, statistical hypothesis tests are performed that compares the predictions of the AI model with the assessments from three dentists.

Enrollment: 31 patients

Sex: All

Ages: 8 to 13 years old

Volunteers: Accepts Healthy Volunteers

Inclusion criteria

• Anterior young permanent teeth

Exclusion criteria

• Anterior young permanent teeth exhibiting disruptions in enamel tissue integrity such as decay
• Hypoplasia, hypomineralization
• Restored and prosthetically treated teeth
• Young permanent teeth located in the posterior region
• Primary teeth

Trial design

Primary purpose: Diagnostic

Allocation: Non-Randomized

Interventional model: Factorial Assignment

Masking: Triple Blind

---

31 participants in 2 patient groups

Deep Learning Models Group

Experimental group

Description:

As artificial intelligence models, DeepLabV3+, Mask R-CNN (Detectron2), YOLOv8, U-Net, Super Vision U-net and U-Net Transformer models, which are state-of-the-art in semantic segmentation, were selected.

Treatment:

Diagnostic Test: Deep Learning Models

The Difference Between The AI Model (U-Net Transformer) and Dentists Group

Active Comparator group

Description:

Using the prior knowledge (α = 0.05, β = 0.2) and an effect size of 0.61, the actual power of the comparison between the AI model (U-Net Transformer) and dentists on 34 test images is at least 80%, which is deemed sufficient. Therefore, randomly selected 35 images on the test dataset were labeled by three dentists without seeing the ground truth and were predicted by the AI model. Then, the intersection over union (IoU) score of these labeled and predicted images were calculated. The IoU score, which computes the ratio between the intersection and the union of two sets, is commonly used to evaluate the accuracy of prediction on semantic segmentation. To confirm clinical feasibility, three t-tests, which evaluates the difference between the means of two variables, were applied to IoU scores of dentists and IoU scores of the AI model and a p value \< .05 was considered statistically significant.

Treatment:

Diagnostic Test: The Difference Between The AI Model and Dentists Group