Extracting the impacted mandibular third molar is one of the routine procedures performed by oral and maxillofacial surgeons [1]. However, the difficulty of extraction varies from simple extraction to cases requiring general anesthesia. Categorizing the difficulty of extractions or estimating the time required has been of interest to oral and maxillofacial surgeons.
MacGregor made the first attempt to establish a model for assessing surgical difficulty [2]. Winter's and Pell & Gregory classification is a classic method [2, 9, 10]. Many studies have found that radiological factors are related to surgical difficulty for both upper and lower third molars, with depth of impaction, distal space available, molar angulation, and root morphology being main variables contributing to the difficulty [11–15]. However, these methods have recently been found to be inappropriate for judging the difficulty of surgery. The Pederson index showed low sensitivity and specificity in predicting the difficulty of surgery for impacted mandibular third molars [16]. Furthermore, the unreliability of radiographs for classification of impacted third molar irrespective of training or experience of the evaluator has been reported [16–18]. Other important variables not calculated by Pederson include maximum mouth opening, age, bone density, body mass index (BMI, kg/m2), and proficiency of the surgeon.
In this study, to measure the difficulty of extracting the mandibular third molar, the length of surgery was used as a target variable. A set of predictor variables were divided into three groups: patient, operator, and radiologic variables. Patient variables included age, gender. BMI, and maximum mouth opening. Radiologic variables included angulation, depth, bone density, morphology of the third molar, and the space between mandibular ramus and mandibular second molar. These radiological factors were not specified as numerical or nominal variables. Radiographs were imported as image data and included into the CNN model.
Methods such as Ridge Regulation (L2 Regularization), Dropout, Early Stopping, and Batch Normalization were applied to the model for normalization. Dropout can prevent the model from overfitting in a way that does not involve some weights in training. Early Stopping is a method of terminating training if the performance of the model for the validation set does not improve any more during the epoch. Batch Normalization is a way to train a model faster and more stably by normalizing the input distribution of each layer. Data augmentation is commonly used to overcome limitations of small data sets that are unique to the medical field. The size of each training dataset was increased from 624 to 8,602 after data augmentation.
The deep learning approach related to the difficulty prediction of third molars uses panoramic radiographs only. De Tobel developed an automated method to assess the degree of development using mandibular third molars on panoramic radiographs [19]. Another study that predicted the difficulty of wisdom tooth extraction using CNN considered only radiological characteristics for the prediction [5]. Unlike other deep learning studies using only panorama or cone beam computed tomography (CBCT), this study is significant in that it is the first study to create a model for predicting the difficulty of extraction of the mandibular third molar by considering clinical data. In this study, a statistically significant positive correlation between the predicted extraction time and real extraction time was observed (Pearson coefficient = 0.8315, p < 0.01).
Renton has reported that age, patient weight, and ethnicity are associated with extraction times [3]. When patients were divided by age, those over 30 years were at a significantly more risk of difficult extractions than younger patients. The difficulty was further increased as patient's age exceeded 50 years (p < 0.05) [3]. However, in the present study, no significant differences in the difficulty of wisdom tooth extraction were found according to age. This might be because age distribution of the study population was biased towards younger patients. A majority of patients included in the study were in their 20s (Fig. 1).
To the best of our knowledge, this is the first study to predict the difficulty of wisdom tooth extraction through artificial intelligence using both panoramic images and clinical data. This study has some limitations. First, the study population of the study was skewed toward the younger age group. The number of subjects in the age group of more than 50 years in the study population was limited. Such age inhomogeneity of this study group might have led to less consideration of the age factor when artificial intelligence was used to predict the difficulty of tooth extraction. If the number of subjects in the entire study group is increased through additional studies in the future with age group of patients uniformly included, a more sophisticated extraction difficulty prediction model is expected to emerge.
In addition, in this study, a panoramic image was used as an image variable. If CBCT data are used for the same prediction model in the future, a more sophisticated model might be obtained.
Experienced dentists and oral and maxillofacial surgeons predict and prepare for the difficulty of tooth extraction by considering various factors such as panorama, CBCT, patient's age, gender, and morphology. However, in the case of novice dentists, the difficulty is often unpredictable. Novice dentists often fail to predict the difficulty of extraction, leading to a very long extraction time or an increase in patient discomfort after surgery. The AI prediction model of this study is an AI model that can predict the difficulty of wisdom tooth extraction by considering both radiographic images and patient's clinical data rather than simply predicting the difficulty through AI using a panoramic image. This can be of great help to novice dentists when predicting the difficulty before wisdom tooth extraction or when deciding whether to refer to a specialist without extraction.