Nantakeeratipat Teerachate, Apisaksirikul Natchapon, Boonrojsaree Boonyaon, Boonkijkullatat Sirapob, Simaphichet Arida
Department of Conservative Dentistry and Prosthodontics, Faculty of Dentistry, Srinakharinwirot University, Bangkok, Thailand.
Faculty of Dentistry, Srinakharinwirot University, Bangkok, Thailand.
Front Dent Med. 2024 Nov 28;5:1507705. doi: 10.3389/fdmed.2024.1507705. eCollection 2024.
To detect dental plaque, manual assessment and plaque-disclosing dyes are commonly used. However, they are time-consuming and prone to human error. This study aims to investigate the feasibility of using Google Cloud's Vertex artificial intelligence (AI) automated machine learning (AutoML) to develop a model for detecting dental plaque levels on permanent teeth using undyed photographic images.
Photographic images of both undyed and corresponding erythrosine solution-dyed upper anterior permanent teeth from 100 dental students were captured using a smartphone camera. All photos were cropped to individual tooth images. Dyed images were analyzed to classify plaque levels based on the percentage of dyed surface area: mild (<30%), moderate (30%-60%), and heavy (>60%) categories. These true labels were used as the ground truth for undyed images. Two AutoML models, a three-class model (mild, moderate, heavy plaque) and a two-class model (acceptable vs. unacceptable plaque), were developed using undyed images in Vertex AI environment. Both models were evaluated based on precision, recall, and F1-score.
The three-class model achieved an average precision of 0.907, with the highest precision (0.983) in the heavy plaque category. Misclassifications were more common in the mild and moderate categories. The two-class acceptable-unacceptable model demonstrated improved performance with an average precision of 0.964 and an F1-score of 0.931.
This study demonstrated the potential of Vertex AI AutoML for non-invasive detection of dental plaque. While the two-class model showed promise for clinical use, further studies with larger datasets are recommended to enhance model generalization and real-world applicability.
为了检测牙菌斑，通常采用人工评估和菌斑显示剂。然而，它们耗时且容易出现人为误差。本研究旨在探讨使用谷歌云的Vertex人工智能（AI）自动化机器学习（AutoML）开发一个模型的可行性，该模型用于使用未染色的照片图像检测恒牙上的牙菌斑水平。
使用智能手机摄像头拍摄100名牙科学生未染色的以及相应的赤藓红溶液染色的上前恒牙的照片图像。所有照片都裁剪为单个牙齿图像。对染色图像进行分析，根据染色表面积的百分比对菌斑水平进行分类：轻度（<30%）、中度（30%-60%）和重度（>60%）类别。这些真实标签被用作未染色图像的基本事实。在Vertex AI环境中使用未染色图像开发了两个AutoML模型，一个三类模型（轻度、中度、重度菌斑）和一个两类模型（可接受与不可接受菌斑）。两个模型均基于精确率、召回率和F1分数进行评估。
三类模型的平均精确率为0.907，在重度菌斑类别中精确率最高（0.983）。在轻度和中度类别中，错误分类更为常见。两类可接受-不可接受模型表现出更好的性能，平均精确率为0.964，F1分数为0.931。
本研究证明了Vertex AI AutoML在无创检测牙菌斑方面的潜力。虽然两类模型显示出临床应用的前景，但建议使用更大的数据集进行进一步研究，以提高模型的泛化能力和实际适用性。
