Artificial intelligence in positioning between mandibular third molar and inferior alveolar nerve on panoramic radiography | Scientific Reports -…

This study evaluated if AI could determine the positional relationship between M3 and IAN based on panoramic radiography regarding whether the two structures were in contact or intimate and whether the IAN was positioned lingually or buccally to M3 when two structures were overlapped. In this situation, determining the exact position was limited and unreliable even for the specialist, as shown in previous studies25,26. However, AI could determine both positions more accurately than OMFS specialists.

Until now, if M3 and IAN overlap on panoramic radiograph, specialists could use the known predictive signs of IAN injury to determine the positional relationship whether the two structures were in contact or intimate. Umar et al. compared the positional relationship between IAN and M3 through panoramic radiography and CBCT. Loss of the radiopaque line and diversion of the canal on panoramic radiographs resulted in tooth and nerve contact in 100% of the cases on CBCT. Darkening of the roots were associated with contact on CBCT in 76.9% of the cases studied27. However, another study reported that the sensitivities and specificities ranged from 14.6 to 68.3% and from 85.5 to 96.9%, respectively, for those three predictive signs1. Datta et al. compared those signs with the clinical findings during surgical removal and found that only 12% of patients with positive radiological signs showed clinical evidence of involvement3. In the present study, we adopted CBCT reading results instead of radiological signs on panoramic radiography to determine the positional relationship so that the AI could determine whether the two structures were in contact or intimate, showing an accuracy of 0.55 to 0.72. Compared to another study1, our deep learning model exhibited similar performance (accuracy 0.87, precision 0.90, recall, 0.96, F1 score 0.93, and AUC 0.82) to determine whether M3 is contacting the IAN or not. This could explain the different model performance depending on the characteristics of the data.

To replace CBCT with analysis of panoramas with AI, information about bucco-lingual positioning was necessary to ensure safe surgical outcomes. It has been reported that the lingual position of the nerve to the tooth has a significantly higher risk of IAN injury compared to other positions28. To the best of our knowledge, no studies have evaluated bucco-lingual positioning through panoramic radiograph because there were no methods to predict this position using one radiograph. Two intraoral radiographs with different angle (vertical tube-shift technique) in the third molar area caused patient discomfort and nausea during placement of the film or sensor of the digital intraoral x-ray devices29 and is difficult to use clinically. Since there was no effective method to discern the position, the accuracy of the specialists was low in this study. On the contrary, the AI showed considerably high accuracy ranges from 67.7 to 80.6% despite the small amount of study data. The course of the IAN predominantly is buccal to the tooth28, and our data revealed a similar situation. However, the total number of cases was small to match the numbers in each group evenly for deep learning. In addition, the lack of total number of cases forced the use of a simple deep learning model with a relatively small number of parameters to be optimized. Therefore, training AI with more data could produce more accurate results and be used more widely in clinical settings.

In this study, bucco-lingual determination (Experiment 2) exhibited superior performance for true contact positioning (Experiment 1). The difference in accuracy between the two experiments seems to be a characteristic of the data rather than a special technical difference. There might be a particular advantage for AI to be recognized in bucco-lingual classification, or that some of the contact classification data might have characteristics that are difficult to distinguish.

There are several studies that have developed Al algorithms that have been able to outmatch specialists in terms of performance and accuracy. AI assistance improved the performance of radiologists in distinguishing coronavirus disease 2019 from pneumonia of other origins in chest CT30. Moreover, the AI system outperformed radiologists in clinically relevant tasks of breast cancer identification on mammography31. In the present study, the AI exhibited much higher accuracy and performance compared to those of OMFS specialists. To determine the positional relationship between M3 and IAN, we performed preliminary tests to determine the most suitable AI model using VGG19, DenseNet, EfficientNet, and ResNet-50. ResNet showed higher AUC in Experiment 2 and comparable AUC in Experiment 1 (Supplemental Tables 13). Therefore, it was chosen as the final AI model.

This study has limitations. First, the absolute size of the training dataset was small. Data augmentation by image modification was used to overcome the limitation of a small sized dataset. Nevertheless, as shown in Table 1, there were cases where training did not proceed robustly. Therefore, the performances of the trained models highly depend on the train-test split. This unsoundness of the trained model, which hinders the clinical utility of AI models for primary determination in practice, can be alleviated by collecting more data and using them for training. Also, the size of deep learning models is an important factor in performance and, in general, a large number of instances are required to train a large size deep neural network without overfitting. Thus, not only collecting more data but also exploiting external datasets from multiple dental centers can be considered to increase the performance of AI models. However, this study is meaningful in that the AI model performed better than experts even under these adverse conditions. Second, the images used in this study were cropped without any prior domain knowledge such as proper size or resolution to include sufficient information to determine true contact or bucco-lingual positional relationship between M3 and IAN. If the domain knowledge is reflected to construct a dataset, the performances of AI models can be highly increased. Third, the use of interpretable AI models32, which can explain the reason for the model prediction, can help to identify the weaknesses of the trained models. The identified weaknesses can be overcome by collecting data that the models have difficulty in classifying. Finally, the various techniques developed in the machine learning society, such as ensemble learning33, self-supervised learning34, and contrastive learning35, can be utilized for further improvement of the performance of our models even in situations where the total number of cases is insufficient as well.

Read more:

Artificial intelligence in positioning between mandibular third molar and inferior alveolar nerve on panoramic radiography | Scientific Reports -...