BackgroundWith recent advances in artificial intelligence, the use of this technology has begun to facilitate comprehensive tissue evaluation and planning of interventions. This study aimed to assess different convolutional neural networks (CNN) in deep learning algorithms to detect keratinized gingiva based on intraoral photos and evaluate the ability of networks to measure keratinized gingiva width.MethodsSix hundred of 1200 photographs taken before and after applying a disclosing agent were used to compare the neural networks in segmenting the keratinized gingiva. Segmentation performances of networks were evaluated using accuracy, intersection over union, and F1 score. Keratinized gingiva width from a reference point was measured from ground truth images and compared with the measurements of clinicians and the DeepLab image that was generated from the ResNet50 model. The effect of measurement operators, phenotype, and jaw on differences in measurements was evaluated by three‐factor mixed‐design analysis of variance (ANOVA).ResultsAmong the compared networks, ResNet50 distinguished keratinized gingiva at the highest accuracy rate of 91.4%. The measurements between deep learning and clinicians were in excellent agreement according to jaw and phenotype. When analyzing the influence of the measurement operators, phenotype, and jaw on the measurements performed according to the ground truth, there were statistically significant differences in measurement operators and jaw (p < 0.05).ConclusionsAutomated keratinized gingiva segmentation with the ResNet50 model might be a feasible method for assisting professionals. The measurement results promise a potentially high performance of the model as it requires less time and experience.PLAIN LANGUAGE SUMMARYWith recent advances in artificial intelligence (AI), it is now possible to use this technology to evaluate tissues and plan medical procedures thoroughly. This study focused on testing different AI models, specifically CNN, to identify and measure a specific type of gum tissue called keratinized gingiva using photos taken inside the mouth. Out of 1200 photos, 600 were used in the study to compare the performance of different CNN in identifying gingival tissue. The accuracy and effectiveness of these models were measured and compared to human clinician ratings. The study found that the ResNet50 model was the most accurate, correctly identifying gingival tissue 91.4% of the time. When the AI model and clinicians' measurements of gum tissue width were compared, the results were very similar, especially when accounting for different jaws and gum structures. The study also analyzed the effect of various factors on the measurements and found significant differences based on who took the measurements and jaw type. In conclusion, using the ResNet50 model to identify and measure gum tissue automatically could be a practical tool for dental professionals, saving time and requiring less expertise.

中文翻译：

---

基于卷积神经网络的口腔角化牙龈检测与测量效率


背景随着人工智能的最新进展，该技术的使用已开始促进全面的组织评估和干预规划。本研究旨在评估深度学习算法中不同的卷积神经网络 (CNN)，以根据口内照片检测角化牙龈，并评估网络测量角化牙龈宽度的能力。方法对使用显示剂之前和之后拍摄的 1200 张照片中的 600 张进行分析用于比较分割角化牙龈的神经网络。使用准确度、并集交集和 F1 分数来评估网络的分割性能。根据地面真实图像测量参考点的角化牙龈宽度，并与临床医生的测量值和 ResNet50 模型生成的 DeepLab 图像进行比较。通过三因素混合设计方差分析（ANOVA）评估测量操作员、表型和颌骨对测量差异的影响。结果在比较的网络中，ResNet50 区分角化牙龈的准确率最高，为 91.4%。根据颌骨和表型，深度学习和临床医生之间的测量结果非常一致。在分析测量操作员、表型和下颌对根据地面实况进行的测量的影响时，测量操作员和下颌存在统计上显着的差异（ p < 0.05）。结论使用 ResNet50 模型进行自动角化牙龈分割可能是协助专业人员的可行方法。测量结果保证了模型的潜在高性能，因为它需要更少的时间和经验。简单语言摘要随着人工智能 (AI) 的最新进展，现在可以使用该技术来评估组织并彻底规划医疗程序。这项研究的重点是测试不同的人工智能模型，特别是 CNN，利用口腔内拍摄的照片来识别和测量一种称为角化牙龈的特定类型的牙龈组织。研究中使用了 1200 张照片中的 600 张来比较不同 CNN 在识别牙龈组织方面的性能。对这些模型的准确性和有效性进行了测量，并与人类临床医生的评分进行比较。研究发现，ResNet50 模型是最准确的，正确识别牙龈组织的正确率为 91.4%。当将人工智能模型和临床医生测量的牙龈组织宽度进行比较时，结果非常相似，特别是在考虑到不同的颌骨和牙龈结构时。该研究还分析了各种因素对测量结果的影响，发现根据测量者和下颌类型存在显着差异。总之，使用 ResNet50 模型自动识别和测量牙龈组织可能成为牙科专业人员的实用工具，可以节省时间并减少专业知识的要求。