Since the outbreak of COVID-19, many hospitals have become overloaded with patients seeking examination, resulting in an imbalance between medical staff and patients. These high concentrations of people in hospital settings not only aggravate the risk of cross-infection among patients, but also stall the public medical system. Consequently, mild and chronic conditions cannot be treated effectively, and eventually develop into serious diseases. Therefore, the use of deep learning to accurately and efficiently analyze X-ray images for diagnostic purposes is crucial in alleviating the pressure on medical institutions during epidemics. The method developed in this study accurately detects dental X-ray lesions, thus enabling patients to self-diagnose dental conditions.

The method proposed in this study employs the YOLOV5 algorithm to detect lesion areas on digital X-ray images and optimize the network model’s parameters. When hospitals and medical professionals collect and label training data, they use image normalization to enhance the images. Consequently, in combination with the network environment, parameters were adjusted into four modules in the YOLOV5 algorithm. In the Input module, Mosaic data enhancement and adaptive anchor box algorithms are used to generate the initial box. The focus component was added to the Backbone module, and a CSP structure was implemented to determine the image features. When the obtained image features are input into the Backbone module, the FPN and PAN structures are used to realize feature fusion. Subsequently, GIOU_Loss function is applied to the Head moudule, and NMS non-maximum suppression is used to generate a regression of results.

The proposed YOLOV5-based neural network yields satisfactory training and testing results. The training algorithm produced a recall rate of 95%, accuracy rate of 95%, and F1 score of 96%. All evaluation criteria are higher than those of the target detection algorithms of SSD and Faster-RCNN (Table 1). The network converges to smoothness after loss is reduced in the training process (Fig. 6), which proves that the network successfully learns the necessary features. Thus, the difference between predicted and real values is very small, which indicates good model performance. The mAP value of network training is 0.985 (Fig. 7), which proves that the network training meets the research requirements. Finally, an observation of the visualized thermodynamic diagram reveals that the network’s region of interest matches the target detection region (Fig. 8).

This study proposes the use of the YOLOV5 algorithm for detecting lesions in dental X-ray images, training and testing on the dataset, modifying the network’s nominal batch size, selecting an appropriate optimizer, adjusting the weight parameters, and modifying the learning rate attenuation strategy. The model’s training results were compared with those of algorithms used in previous studies. Finally, the effect of feature extraction was analyzed after the thermodynamic diagram was visualized. The experimental results show that the algorithm model detects lesion areas with an accuracy rate of more than 95%, making it an effective autonomous diagnostic tool for patients.