基于自监督迁移学习与ResNet50的牙周炎曲面断层片分类研究

Objective To construct an automatic classification model of moderate to severe periodontitis based on selfsupervised transfer learning and ResNet50, and to analyze the feasibility of its clinical application. Methods A total of 93 panoramic radiographs of healthy control group and 93 panoramic radiographs of moderate to severe periodontitis group were included in this study. A two-stage transfer learning framework was constructed and its clinical value was evaluated after data enhancement with online data. Results With the increase of rounds, the training loss decreased rapidly and stabilized to 0.2～0.3. The verification loss first decreased and then stabilized without significant divergence. The training accuracy continued to increase to more than 95%, and the internal validation accuracy continued to increase and remained stable at 82%～86%. In the internal validation set(38 images), the accuracy, sensitivity, specificity and AUC were 84.21%, 73.68%, 94.74% and 0.89(95%CI: 0.78～0.96), respectively. The precision and recall of the health category were 0.78 and 0.95, and those of the periodontitis category were 0.93 and 0.74. The high specificity of the model can effectively exclude healthy individuals without bone resorption, reducing the workload of basic screening for physicians. However, insufficient sensitivity may lead to missed diagnosis of periodontitis, making it only an auxiliary screening tool and not a sole basis for diagnosis. Conclusion The classification of periodontitis by panoramic radiographs can be applied in clinical practice, but its value still needs to be further verified.