Ai Dental Diagnosis

[Jeremias Resendez]

Thankyou for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

Radiological examination is an essential part of patient management in dentistry. It is frequently used to supplement and aid clinical diagnosis of pathology related to teeth and adjacent structures1,2,3,4. Cone-beam computed tomography (CBCT) was proposed for maxillofacial imaging5,6 during the last decade and is now becoming increasingly popular for such use. It offers distinct advantages including lower radiation doses, compared to medical computed tomography (CT), and the potential of importing and exporting individualized, digital imaging and communications in medicine (DICOM) and overlap-free reconstructed data for other applications4,5,6,7. CBCT can supply high-resolution three-dimensional (3D) images without distortion and superimposition of bone and other dental structures that can be seen in conventional radiography8,9.

Several studies have compared the diagnostic accuracy of CBCT with conventional and digital panoramic and periapical radiography10,11,12,13,14. CBCT has been shown to significantly increase the detection rate of tooth root canal spaces and periapical areas for the evaluation of dental infection and pathology compared to conventional imaging15,16,17. This suggests that CBCT enhances the recognition of periapical bone lesions and offers improved diagnostic accuracy, treatment planning, and thus prognostic outcomes. These and other possibilities along with increasing access to CBCT imaging for dentists are allowing the transition from 2 to 3D imaging in everyday dental practice10,11,12,13,14,15,16,17.

Previous studies revealed that CBCT has a wide application in the field of dentistry18,19,20. However, there is no standard curriculum or protocol for providing training regarding CBCT. The current status of awareness and knowledge concerning CBCT amongst dental practitioners is not known precisely21.

There have been limited works in the literature studying the knowledge and attitude of dentists toward advanced dentomaxillofacial imaging. The literature showed that there is a lack of knowledge regarding CBCT22,23. It should also be pointed out that studying CBCT should take more time in dental school curricula21. Reddy et al.23 in their work defined very low awareness amongst the dentists regarding CBCT applications, which can be interpreted as a lack of experience in that area. Thus, computer-aided systems have been developed to assist in medical and dental imaging diagnosis24,25,26,27.

Along with the expansion of CBCT, radiation-related effects of CBCT imaging raise concerns about its use in dentistry. CBCT is associated with a higher radiation dose compared to panoramic and intraoral imaging, but a lower dose compared to conventional tomography28,29,30. Therefore radiation risk assessment is in place. The effective dose recommended by ICRP (International Commission on Radiological Protection) should be kept accordingly to the principles of ALARA (As Low as Reasonably Achievable) and ALADA (As Low as Diagnostically Acceptable). It should also be stated that the necessity of CBCT scanning must be indication-oriented and patient specific29.

Convolutional Neural Networks (CNN) are most commonly used for object detection and segmentation. There are several studies available with deep-learning methods, including CNNs, to assist clinicians in dentistry. CNNs are used clinically for, apical lesion detection31, detection of root fractures32, detection of periodontal disease33, cystic lesions34, caries detection35, staging of lower third molar development36, tooth detection37,38, diagnosis of jaw lesion39, and other pathologies detection40. Artificial intelligence (AI) provides an added value and decision support tool for such medical imaging.

There are a few critical success factors to measure the gap between actual performance and expected achievement. AI systems must apply to real-world situations and be designed for clinical evaluation and deployment. Furthermore, an important part of the development and integration of these AI systems is that their functionality (ease of use, speed, and accuracy) reaches or exceeds the clinicians' expertise and expectations.

In this study, a novel AI system, which is based on deep learning methods, was tested for diagnostic capabilities. Firstly, the clinical performance, accuracy, and time required for diagnosis were evaluated. The real-time performance of CBCT imaging was evaluated on the diagnosis of anatomical landmarks and pathologies. Secondly, its clinical effectiveness and safety were tested when used by dentists in a clinical setting. The null hypothesis of this study was that there is no significant difference between aided and unaided groups using the proposed AI system (Diagnocat) for CBCT imaging.

The results of the AI evaluation are shown in Tables 2 and 3. Table 2 shows the overall sensitivity and specificity for the system and dentomaxillofacial radiology examiners. Outcome counting for Table 4 was summarized over the case, tooth, and condition, whilst grouped by the participants. Both sensitivity and specificity were recorded as higher for human examiners. Overall sensitivity values for human examiners ranged between 0.9318 and 0.9438 while the value for this AI system was 0.9239. Overall specificity values for ground truth examiners were between 0.9899 and 0.9946 while the value for this AI system was 0.9899. Table 3 shows sensitivity and specificity values for the system given per condition. The results of specificity values were high, with the lowest being 0.94 when determining missing tooth. Sensitivity values were condition-dependent, with the lowest values being around 0.7 for some difficult or subjective conditions such as endodontic treatment (missed canal, short filling, voids in root filling), and signs of dental caries (complex to diagnose using the CBCT). Notably, Diagnocat struggled to detect very rare anatomical configurations of the tooth e.g., 5 canals or 4 roots. Finally, a rare subtype of the periapical lesion, periapical radiopacity, did not register in the dataset. Yet, this subtype currently is not claimed as a diagnostic capability of the Diagnocat system.

The integration of AI into healthcare has dramatically accelerated in the past decade. The use of deep learning advanced almost synchronously in both medical and dental fields9. Previous studies in dentistry focused on image-processing algorithms to achieve high-accuracy classification and segmentation in dental radiographs. They used mathematical morphology, active contour models, level-set methods, Fourier descriptors, textures, Bayesian techniques, linear models, or binary support vector machines27,41. However, image components are usually obtained manually using these image-enhancement algorithms. The deep learning method used in this AI system (Diagnocat) yielded fairer outcomes by automatically obtaining image features. The objects detected in an image are classified into a pretrained network without preliminary diagnostics, as a result of processes such as filtering and subdivision. With its direct problem-solving ability, deep learning is used extensively in medicine. Deep learning methods using CNNs are a cornerstone of medical image analysis42. Such methods have been preferred in AI studies in dental radiology as well. Tooth detection, identification, and numeration are the first diagnostic steps in dental radiography. Image-processing algorithms have been developed with classification and segmentation in dental radiographs using mathematical morphology, active contour, or level-set methods. A previous study used Bayesian classification for generating an automated dental identification system to classify and identify teeth in bitewing radiographs25. Similarly, another study recommended a tooth classification and numbering system to efficiently segment, classify, and number teeth using an image enhancement technique in bitewing radiographs43. Tooth detection and numbering have been researched intensively during the last few decades mainly using threshold and region-based techniques. CNN as a popular deep learning method has been used to detect and number teeth as well. It was also emphasized that localization of teeth is important for dental image applications, similarly to our results44. In this paper, the authors suggested an original teeth localization technique for periapical radiographs using oriented tooth detection using a CNN. The results of this study showed that the proposed method is effective to localize teeth successfully.

Similar CBCT studies were performed and reported in the literature. One of them considered automatic teeth classifying system using 7 types of axial slice CBCT images using a CNN37. The authors concluded that a 7-tooth type classification system can be used efficiently for automatic dental charting37. Another study also described a CNN model modified with AlexNet architecture for tooth detection in panoramic radiographs38. This study defined mouth gap detection that showed the possible placement of teeth for preprocessing steps. It was concluded that this model could be efficiently used for the detection of teeth. These findings are in line with our study. Another study45 used the sea mask region-based CNN method with transfer learning strategies while another46 used a fully deep learning mask region-based convolutional neural network (R-CNN) method implemented through a fine-tuning process for automated tooth segmentation. This technique showed high performance for automatic teeth segmentation on panoramic radiographs. Similarly, one more study also used the state-of-the-art Faster R-CNN model of tooth detection and numbering47. A recently published paper proposed a deep learning CNN model with a VGG16 network structure for the teeth detection and classification of periapical radiographs27. The CNN method is similar to our study, which can also interfere with deep learning methods and can be used for both training and transfer learning.

3a8082e126