Elsevier

Displays

Volume 71, January 2022, 102119
Displays

Artificial intelligence assisted identification of therapy history from periapical films for dental root canal

https://doi.org/10.1016/j.displa.2021.102119Get rights and content

Highlights

  • According to the characteristics of tooth morphology, we have proposed a method for extracting Region of Interest (ROI) tooth root canal.

  • Three machine learning methods have been utilized for training and testing, respectively, with and without our proposed ROI extraction operation.

  • The data enhancement method improves the generalization ability of the model.

  • The result proved that our proposed method of ROI extraction is effective for the artificial intelligence assisted diagnosis on the history of root canal therapy.

Abstract

In this work, we aim to develop and validate an AI-assisted method for identifying the history of root canal therapy by using periapical films. First, we propose a pre-processing method to extract the regions of interest (ROI) containing the root canals. Then, in order to improve the generalization ability, data augmentation is adopted to expand the dataset. Finally, three machine learning methods, including SIFT-SVM, CNN, and transfer learning are used. All the models are validated based on the receiving operating characteristic (ROC) curve analysis. The accuracies for the three machine learning methods are above 95%. The AUC, sensitivity, and specificity of the best model are also presented and analyzed.

Introduction

The auxiliary diagnosis of X-ray images based on artificial intelligence (AI) is a research hotspot and an imperative development trend in clinical medicine. Currently, there are various types of medical images, such as magnetic resonance imaging (MRI), computed tomography (CT), and X-rays. The auxiliary diagnosis based on AI has been effectively applied to various medical image classification and segmentation problems, such as liver tumor image segmentation, kidney tumor image segmentation, and survival judgment of patients with medulloblastoma [1], [2], [3]. At the same time, AI is also used to enhance the quality of medical images [4]. The auxiliary diagnosis based on AI is of great academic significance, social benefit, and economic value, as it provides a means for diagnosis, developing telemedicine, and improving the quality of medical therapy in remote and poor neighborhoods. Please note that the AI-assisted diagnosis of oral X-ray images is still in its infancy. As a result, there are very few related works presented in literature. Additionally, the works presented in literature generally focus on dental cavities, periodontitis, and wisdom tooth. To the best of our knowledge, there are no solutions based on AI-assisted diagnosis that address root canal therapy. Although the representation of the state after root canal therapy using an X-ray film is rather obvious, and easy to analyze, there is a certain degree of subjectivity in some special cases resulting in misdiagnosis. Therefore, the AI-assisted diagnosis of root canal therapy is an essential technique in clinical medicine.

The pre-processing methods used for dental caries, periodontitis, wisdom teeth and other dental diseases are too simple. It is noteworthy that the dental lesions and post-therapy features are usually reflected in X-ray films. If the corresponding ROI is not extracted, the key features of the relevant dental diseases are not highlighted effectively. Consequently, the trained model is unresponsive to the ROI of the corresponding dental diseases and is not robust, thereby affecting the accuracy of the final diagnosis. In the pre-processing stage, the features of the corresponding diseases are extracted or highlighted for obtaining the training samples. This greatly improves the accuracy of overall diagnosis.

Considering the problems existing in the current research on AI-assisted diagnosis of dental diseases, this work proposes an ROI extraction method of root canal. The data used in this work is enhanced based on the samples of root canal therapy history obtained by the professional doctors. The main idea is to obtain the gingival line and segment the upper and lower images, discard the tooth images above the gingival, and extract the teeth inside the gingival as the final samples. We use three methods, i.e., SIFT-SVM, CNN, and transfer learning to train and test the samples. The overall workflow is shown in Fig. 1.

In Table 1, we present the studies related to the AI-assisted diagnosis of the existing oral X-ray images, the sample pre-processing method used in this work, and machine learning methods.

Section snippets

Preliminaries

The flow chart of the ROI extraction of root canal is shown in Fig. 2. There are two main reasons of interference in the AI-assisted diagnosis of the history of root canal therapy, including teeth fillings and teeth crowns. In these two types of teeth samples, pixels with high brightness (white pixels) are concentrated, similar to positive root canal samples (the teeth samples with a history of root canal therapy). However, these two types of teeth belong to negative root canal samples (the

Samples

The datasets used in this work are obtained from the Peking University Hospital of Stomatology. The selection principles considered during the data collection are as follows:

  • 1.

    The quality of each periapical film should satisfy the requirements of clinical routine diagnosis and therapy.

  • 2.

    All the patients should be older than 18 years.

  • 3.

    The photography regions considered in the periapical films should have normal permanent dentitions. The individuals with extra teeth, malformed teeth, or severe

Learning algorithm

There are few works on the AI-assisted diagnosis for teeth, researchers worked on obtaining the suitable algorithms for different tasks and samples. The purpose of this work is to provide more references for clinical applications, so, three algorithms with great differences in principle were selected. The most important difference among SIFT-SVM, CNN (VGG16), and transfer learning was the difference in the feature extraction. SIFT-SVM extracted the features by looking for SIFT feature points,

Experiments and results

In this work, we use SIFT-SVM, CNN (VGG16), and transfer learning (VGG16, VGG19, and ResNet50) are with and without ROI feature extraction. The hardware platform used in this paper is Nvidia Geforce RTX 2060 GPU. Opencv software framework was used in the SIFT SVM experiment, while CNN and transfer learning experiment was carried out on the Windows platform based on Tensorflow2.0 framework.

Conclusions

In this work, the AI-assisted diagnosis is conducted for the history of root canal therapy by using the oral X-ray images. An extraction method suitable for root canal ROI region is proposed. Based on three machine learning methods, the training and testing are performed with and without using ROI extraction. The results show that after ROI extraction, the overall accuracy is improved to a certain extent, and the robustness of the model also increases in terms of the convergence rate of ROC and

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgements

This work is supported by the Beijing Natural Science Foundation (4192047), and the Fundamental Research Funds for the Central Universities (2019JBM345).

References (16)

There are more references available in the full text version of this article.

Cited by (7)

  • Artificial Intelligence Algorithms and Techniques for Dentistry

    2023, 2023 1st International Conference on Cognitive Computing and Engineering Education, ICCCEE 2023
View all citing articles on Scopus
View full text