Dental caries is a common oral disease affecting the hard tissues of teeth, typically resulting from bacterial infections. This in turn leads to chronic damage. It is prevalent during childhood with symptoms such as the discoloration, deformation, and structural deterioration of teeth. According to Chinas fourth epidemiological survey, 71.9% of five-year-olds experience dental caries in deciduous teeth, and 34.5% of 12-year-olds experience it in permanent teeth. Hence, it is a widespread oral disease among children. Challenges, such as the lack of cooperation from children during tooth brushing and oral examinations, a scarcity of pediatric dentists, and low awareness among parents and pediatricians, often result in the delayed diagnosis and treatment at early stages. Furthermore, variability in diagnostic skills among doctors can result in different diagnoses for the same patient. Therefore, aiding in the identification of dental caries and enhancing diagnostic accuracy are essential for effective clinical diagnosis. Although deep learning-based object detection algorithms have made some progress in detecting dental diseases, they still do not adequately meet the needs for accuracy and speed in diagnosis and identification. Additionally, since detection typically relies on professional medical imaging, it prevents patients from performing self-checks using more accessible devices such as smartphones.

To address these issues, in this study, a dataset comprising oral images acquired using mobile devices, such as smartphones, was created. Compared to the uniform features of oral images captured with professional equipment, these images inevitably varied in aspects such as lighting intensity and shooting angle, making it challenging for the target detection model to fit during training. To overcome these difficulties, in this study, a caries detection algorithm was proposed based on multiscale path aggregation and robust data augmentation methods were proposed. First, architecture of a path aggregation network (PANet) was adopted in the network to enhance the models ability to extract and integrate semantic information at different levels of the image. Second, a feature extraction module that combines cross-stage partial connections and residual connections was used to effectively strengthen the models feature extraction capabilities. Subsequently, a single-stage detection head was employed to predict the detection box information directly from feature maps of different scales, which maintained multiscale target detection capabilities while improving the models detection speed. Finally, during the training process, data augmentation methods, such as cropping, stitching, affine transformations, and random flipping, were used to minimize the impact of image differences on the performance of the model.

The results of CR-PANet on the caries dataset validation set indicate that CR-PANet realizes an mAP@50 of 88.2%, mAP@50-LQ of 84.6%, and FPS of 169, meeting the requirements for precise real-time detection. Compared to other commonly used single-stage target detection models, CR-PANet is slightly weaker in terms of computational speed and parameter quantity. However, it shows improvement in precision, recall, and mAP@50. Given the multiscale path aggregation architecture of the CR-PANet model, it maintains sensitivity to targets of different sizes while reducing the number of parameters and computational load, outperforming dual-stage target detection models in performance (Table 1). The CR-PANet can determine the positions of detection boxes with very low bias and identify target categories with high confidence. Moreover, even on low-quality images with issues such as overexposure, darkness, blurriness, or occlusion, CR-PANet still realizes excellent detection results with higher accuracy than that obtained by the other comparative models (Fig. 8?9). The ablation experiments were conducted in four groups. The first group used PANet as the baseline model, with an mAP@50 of 73.4%, a precision of 79.5%, and a recall of 69.2%. The second group replaced CSPBlock in the baseline network with CRBlock, which increased the mAP@50 by 5.8 percentage points. The third group replaced the dual-stage detection head used in the baseline model with the new detection head, resulting in a 1.4 percentage points increase in mAP@50. The fourth group applied new data-augmentation methods to the training data, significantly improving all four indicators (Table 2). After incorporating the feature extraction module CRBlock, CR-PANet converged faster, and the training process became more stable. With the addition of the detection head, and data augmentation strategies, the detection performance of the model is further enhanced (Fig. 10). By comparing the detection results of CR-PANet on the three test images with real annotations and Grad-CAM, it is evident that CR-PANet can detect almost all dental disease areas and accurately determine their positions and ranges, demonstrating its excellent ability to extract semantic information and retain positional information (Fig. 11).

In this study, CR-PANet target detection model is proposed. The model is designed for the automatic detection of dental disease areas in oral images acquired via mobile devices, such as smartphones. Initially, the model incorporates the multiscale path aggregation architecture of PANet, enhancing its ability to detect targets of varying sizes and positions. Subsequently, cross-level and residual connection structures are combined to propose the feature extraction module CRBlock and detection head. This integration not only reduces the number of parameters and computational load but also improves the feature extraction capabilities. Furthermore, it stabilizes the gradient flow during training. Finally, various data augmentation strategies are employed to diversify the dataset images, thereby improving the generalization ability and robustness of the model. The experimental results show that CR-PANet realizes an mAP@50 of 88.2% on the validation set, with precision and recall rate of 98.9% and 89.0%, respectively. Even on a low-quality dataset, mAP@50 realizes 84.6% accuracy, which is a significant improvement over other target detection models. Additionally, ablation experiments reveal that both CRBlock and detection head significantly enhance the feature extraction capabilities of the model, and data augmentation strategies further improve its generalization ability. In summary, the CR-PANet algorithm enhances the detection of dental disease areas in oral images captured via mobile devices, meeting the requirements for real-time detection. Their accuracy and speed are sufficient for the patients to conduct self-checks. Future research should focus on expanding the dataset, refining the granularity of the dental disease types, and detecting a broader range of dental diseases.