With the continuous development of artificial intelligence, Convolutional Neural Networks (CNN) are gradually being applied in medicine. CNN consists of an input layer, convolutional layer, activation function layer, pooling layer, and fully connected layer. The input layer receives raw image data, which is processed through the convolutional layer to extract image information and form feature maps. The activation function layer maps the inputs of neurons to outputs, while the pooling layer reduces the data in the feature maps to decrease computational load and adapt to image misalignment. Finally, the fully connected layer outputs classification of the features. CNN plays an essential role in interpreting medical imaging without subjective evaluation and has become a mainstream recognition model in the field of computer vision. It is mainly applied in image-related tasks, including recognition, classification, segmentation, labeling, and generation, and can automatically extract complex data through extensive training to accomplish target tasks. This article summarizes the research progress of CNN applications in oral diseases.

1.1 CNN and the Prediction of Oral Diseases CNN can be utilized for the prediction of oral diseases, providing better assistance in treatment. Ngnamsie Njimbouom et al. used a multimodal data-based model to predict caries, achieving an accuracy of 90%. In clinical treatment, patients often prefer to retain their teeth rather than undergo extraction. To assist clinicians in formulating acceptable treatment plans, Lee et al. used 598 periapical X-rays of single-rooted premolars and employed a CNN model to predict the outcomes of pulp treatment, demonstrating its utility in aiding clinical decision-making related to pulp treatment. This model serves as an effective method for assessing the treatment outcomes of pulp diseases, showing promising applications in the diagnosis and treatment of pulp diseases. CNN can also be used to predict the progression risk, survival rate, and metastasis probability of patients with oral cancer (OC). A CNN-based risk stratification model for oral mucosa, after training, can identify abnormal morphological features of oral epithelium, using HE-stained pathological biopsy slices of oral leukoplakia to predict the risk of OC occurrence and screen patients with a lower risk of cancer progression. Fujima et al. utilized a CNN model to identify the heterogeneity of fluorodeoxyglucose uptake within lesions, achieving an accuracy of 80% in predicting the survival rate of OC patients. Shaban et al. trained a new CNN model to quantify tumor-infiltrating lymphocytes in whole-slide digital slices of OC patients, achieving an accuracy of 96.31%, which not only allows for more accurate differentiation of tumor and lymphocyte regions but also demonstrates that the presence of tumor-infiltrating lymphocytes can predict the prognosis of patients’ disease-free survival. Tomita et al. investigated the value of CNN in distinguishing metastatic from non-metastatic cervical lymph nodes in enhanced CT, showing significantly better diagnostic performance than radiologists’ assessments. During orthodontic and implant treatments, operations involving tooth extraction and implant placement require careful handling, especially in osteoporotic patients (particularly those on anti-resorptive medications). Nakamoto et al. constructed a new system to predict osteoporosis and prevent fractures. When osteoporotic patients are treated with bisphosphonates or denosumab, some scholars used CNN models to automatically classify trabecular bone in patients with medication-related osteonecrosis, achieving an accuracy of 96%, which can assist clinicians in identifying high-risk areas for osteonecrosis before performing procedures involving exposed bone (such as tooth extraction or implant placement).

1.2 CNN and the Diagnosis of Oral Diseases CNN has various applications in the diagnosis of oral diseases. In the field of dental pulp diseases, there have been numerous studies on machine detection and classification of caries. Bayraktar and Ayan employed CNN to analyze the imaging dataset of bitewing radiographs, improving the diagnostic accuracy for interproximal carious lesions. Near-infrared transillumination does not produce ionizing radiation and is based on fiber optic technology, where the device emits light onto the teeth. The carious dental tissue absorbs more light than healthy tissue, creating a distinct contrast. Schwendicke et al. applied CNN for caries detection based on near-infrared transillumination imaging, enabling early intervention and treatment for caries. Some studies have also adopted multi-level methods to refine target detection when sufficient imaging datasets are available, employing CNN for caries detection in oral imaging data. Hu et al. conducted automatic diagnosis of root fractures on cone-beam computed tomography (CBCT) images, achieving high accuracy, with an accuracy rate of 97.8% in the manual selection group. Qian Jun et al. utilized artificial intelligence methods to detect, identify, and segment chronic apical periodontitis, achieving a Dice coefficient of 95.93%. These studies demonstrate that CNN shows good efficacy in the diagnosis of dental pulp diseases. In oral and maxillofacial surgery, CNN has strong abilities in identifying the mandibular canal, cysts, and OC. Liu et al. automatically detected the mandibular third molar and mandibular canal based on a U-Nets model and classified their relationship using a ResNet-34 model (average sensitivity of 90.2%, average specificity of 95.0%, average accuracy of 93.3%), significantly reducing the time required. Mohanty et al. used whole-slide digital slices and deep learning methods for histopathological analysis, achieving a classification accuracy of 97% for odontogenic keratocysts. Currently, diagnosing and predicting the prognosis of odontogenic keratocysts using pathology is an emerging research area. Early detection and diagnosis of OC are crucial for disease prognosis. Optical coherence tomography, with its high resolution and non-invasive advantages, can automatically identify OC in scanned images using CNN, achieving an accuracy of 96.76%. Panigrahi et al. applied CNN to classify OC histopathological images, achieving an accuracy of 96.6%. Genomics, proteomics, and multi-omics have also been utilized in the study of OC detection. In periodontology, periodontitis, as a chronic non-communicable disease, often leads to symptoms such as periodontal pus, tooth mobility, and even tooth loss, where early prevention and diagnosis are key to improving treatment outcomes. Chang et al. applied CNN for automatic staging of periodontitis on panoramic radiographs, simplifying the bone destruction patterns of periodontitis and achieving high detection performance for periodontal bone levels. Kim et al. developed a CNN automatic diagnostic system that can not only detect the degree of alveolar bone resorption on panoramic radiographs but also identify the numbered locations of affected teeth, showing that this system can enhance the diagnostic efficiency of alveolar bone resorption. In orthodontics, several studies have employed CNN for analysis of cephalometric radiographs and CBCT measurements to assist in diagnosing orthodontic cases. In oral implantology, the incidence of peri-implantitis is high, which, if not treated promptly, may lead to implant loosening or loss, making early diagnosis crucial. Lee et al. established a CNN model for detecting peri-implantitis, achieving 100% specificity and precision, but requiring more images from other implant systems to enhance its learning performance for clinical application. In other diseases (such as temporomandibular joint disorders), magnetic resonance imaging (MRI) is currently the most accurate method for diagnosing temporomandibular joint disc displacement. Lee et al. constructed an automatic diagnostic model for MRI temporomandibular joint disc displacement, achieving an accuracy of 77%. Kao et al. applied CNN to automatically detect temporomandibular joint disc displacement, achieving an accuracy of 85%.

CNN has advantages in processing databases, with strong capabilities for rapid image diagnosis and high diagnostic accuracy, assisting clinicians in improving work efficiency and achieving effective patient care. Although there are certain limitations in the application of CNN in oral diseases, its use is a trend in the development of the discipline, and the application of CNN models in clinical practice is crucial. Future research can focus on the timing of oral disease formation to address the shortcomings of detecting disease activity in clinical practice, further constructing a richer database, and developing patient-centered artificial intelligence systems that integrate data from individuals and environments to create personalized dental care. Efficient utilization of CNN in artificial intelligence to establish learning models can not only assist clinicians in decision-making, improve work efficiency, reduce medical risks and costs but also promote the development of smart healthcare.