The present clinical study obtained approval from the Ethics Committee of Zhujiang Hospital, Southern Medical University (2023-KY-050-01). The project was registered and approved in Medical Research Registration System (medicalresearch.org.cn) of National Health Security Information Platform (MR-44-23-015681).

We prospectively enrolled consecutive patients visiting the endodontics clinics (J. Z. and F. Z.) at Zhujiang Hospital, Southern Medical University, from April to June 2023. Patients under 18 years of age, with mouth opening ≤ 3.5 cm, or unable to complete additional examinations (e.g., X-ray) for diagnosis purpose were excluded. The written informed consent was obtained from each volunteer prior to participation.

A total of 4361 teeth in 191 participants were recorded. Among these participants, 126 (71%) were women and 18 (29%) were men. The average age was 46 years (range 23–86 years). Of the 4361 teeth, 1517 were anterior teeth, 1518 were premolars, and 1326 were molars.

Caries detection was achieved with a commercially available device (Aiyakankan, Aicreate, Zhuhai, China). Briefly, the architecture employs MobileNet-v3 as the backbone network for feature extraction and combines it with a 5-layer U-net for semantic segmentation of teeth. The entire model-building process used a dataset consisting of over 150,000 professionally visually annotated images. Of these, 15% of the images were set aside for testing. Image augmentation techniques are applied to enhance network robustness, including random adjustments in translation, cropping, rotation, brightness, and chromaticity. Images are resized to 512 × 512 and normalized before training using the PyTorch framework. MobileNet-v3 is chosen for its compact size, low computational requirements, and compatibility with mobile devices. Depthwise separable convolutions, residual structures from ResNet, and attention mechanisms are incorporated for accurate predictions. Batch normalization and dropout ensure training stability. The learning rate follows a dynamic decay algorithm with a factor of 0.99. MobileNet-v3 provides 5 feature outputs that are combined with U-net using a 5-layer upsampling and concatenation approach for pixel-level classification. The training process includes multiple iterations, each building upon the previous checkpoint. Each iteration involves 100 epochs, starting with a batch size of 8 and reducing it to 2. The loss function is cross-entropy, and optimization is performed using the Adam optimizer.

Image acquisition was performed using an A9X intraoral camera (Aicreate), maintaining a distance of 20–50 mm from the dental surfaces. Images were captured at a resolution of 1920 × 1080 pixels and stored in JPEG format. Three distinct views—incisor/occlusal, vestibular/buccal, and palatal/lingual—were obtained for each tooth. Images exhibiting inadequate quality, such as those out of focus, fogged, or incomplete, were excluded and retaken. A tooth was classified as negative only if the AI detected no evidence of carious lesions across all three views.

The intraoral camera was operated through a provided computer system, which was included as part of the A9X package from the manufacturer Aicreate. The system specifications included a monitor resolution of 1920 × 1080 pixels, a 60 Hz refresh rate, an Intel i7-3537U CPU, 8GB RAM, a 256GB SSD, and the Windows 10 operating system.

One of the authors (J. Z.) conducted all image acquisitions and quality assessment. After photographing, J. Z. and F. Z. conducted diagnostic confirmation and subsequent treatment in accordance with clinical routine. The observers did not modify the images and were unaware of the AI’s assessments until the completion of all experiments. The AI’s findings were strictly employed for research purposes and did not factor into any clinical decisions.

For each patient enrolled in the study, carious lesions were evaluated according to the WHO standard (Oral Health Surveys: Basic Methods, 5th edition) [9] by J. Z. and F. Z. (each with 5 years of clinical practice) independently. In cases of differing opinions, a more experienced doctor (X. S., with over 20 years of clinical practice) makes the final decision, providing rationale until the three doctors reach a consensus diagnosis.

Data were analyzed using GraphPad Prism 8.3 (GraphPad Software). The values of true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN) were determined by comparing them to the clinical diagnosis, which served as the reference standard. Sensitivity (SE), specificity (SP), positive predictive value (PPV), and negative predictive value (NPV) were calculated for different types of caries. Accuracy (ACC) was defined as ACC = (TN + TP) / (TN + TP + FN + FP). F1 Score were calculated as F1 = 2TP / (2TP + FP + FN).