Khorshidi Faezeh, Esmaeilyfard Rasool, Paknahad Maryam
Computer Engineering and Information Technology Department, Shiraz University of Technology, Shiraz, Iran.
Oral and Dental Disease Research Center, Oral and Maxillofacial Radiology Department, Dental School, Shiraz University of Medical Sciences, Shiraz, Iran.
Saudi Dent J. 2024 Dec;36(12):1582-1587. doi: 10.1016/j.sdentj.2024.11.007. Epub 2024 Nov 26.
Forecasting the complexity of extracting mandibular third molars is crucial for selecting appropriate surgical methods and minimizing postoperative complications. This study aims to develop an AI-driven predictive model using CBCT reports, focusing specifically on predicting the difficulty of mandibular third molar extraction.
We conducted a retrospective study involving 738 CBCT reports of mandibular third molars. The data was divided into a training set consisting of 556 reports and a validation set containing 182 reports. The study involved two main steps: pre-processing and processing of the textual data. During pre-processing, the reports were cleaned and standardized. In the processing phase, a rule-based NLP algorithm was employed to identify relevant features such as angulation, number of roots, root curvature, and root-nerve canal relationship. These features were utilized for the training of a deep learning neural network to classify the extraction difficulty into four categories: easy, slightly difficult, moderately difficult, and very difficult.
The classification model achieved an accuracy of 95% in both the training and validation sets. Precision, recall, and F1-score metrics were calculated, yielding promising results with precision and recall values of 0.97 and 0.95 for the training set, and 0.97 and 0.89 for the validation set, respectively.
The study demonstrated the high reliability of AI-based models to forecast the complexity of the mandibular third molar extractions from CBCT reports. The results indicate that AI-driven models can accurately predict extraction difficulty, thereby aiding clinicians in making informed decisions and potentially improving patient outcomes.
预测下颌第三磨牙拔除的复杂性对于选择合适的手术方法和尽量减少术后并发症至关重要。本研究旨在利用CBCT报告开发一种人工智能驱动的预测模型，特别关注预测下颌第三磨牙拔除的难度。
我们进行了一项回顾性研究，纳入了738份下颌第三磨牙的CBCT报告。数据被分为由556份报告组成的训练集和包含182份报告的验证集。该研究包括两个主要步骤：文本数据的预处理和处理。在预处理过程中，对报告进行了清理和标准化。在处理阶段，采用基于规则的自然语言处理算法来识别相关特征，如角度、牙根数量、牙根曲率和牙根与神经管的关系。这些特征被用于训练深度学习神经网络，以将拔除难度分为四类：容易、稍难、中等难度和非常困难。
分类模型在训练集和验证集中的准确率均达到95%。计算了精确率、召回率和F1分数指标，结果令人满意，训练集的精确率和召回率分别为0.97和0.95，验证集的精确率和召回率分别为0.97和0.89。
该研究证明了基于人工智能的模型从CBCT报告预测下颌第三磨牙拔除复杂性的高可靠性。结果表明，人工智能驱动的模型可以准确预测拔除难度，从而帮助临床医生做出明智的决策，并可能改善患者的治疗结果。
