ADOPTION OF X-RAY IMAGE BONE DENSITY INTELLIGENT RECOGNITION IN EARLY DIAGNOSIS OF PERIODONTAL DISEASE

Abstract

This study aimed to assess the application value of the optimized ResNet50 model in the early diagnosis of periodontal disease using X-ray images. A total of 1000 X-ray images from the medical information mart for the intensive care chest X-ray (MIMIC-CXR) database were utilized, comprising 400 normal images and 600 images indicative of early periodontal disease, which were subsequently divided into training, validation, and testing sets. Data augmentation techniques were employed to enhance the model’s performance, while the impacts of varying learning rates (LRs, 0.0001, 0.001, 0.01) and batch sizes (BSs, 16, 32, 64) on the model’s accuracy, precision, recall, and F1-score were evaluated. Furthermore, the optimized ResNet50 model was compared with pain assessment via image network (PAINet), faster region-based convolutional neural network (Faster R-CNN), and an improved ResNet-18 model, using a comparative dataset of 5000 clinical X-ray images (including 3000 healthy images and 2000 images of periodontal disease). The model’s interpretability was analyzed using gradient-weighted class activation mapping (Grad-CAM) techniques. The results indicate that data augmentation significantly enhanced the model’s performance, achieving optimal results with an LR of 0.001 and a BS of 32. The optimized ResNet50 model demonstrated statistically superior accuracy and AUC values compared to PAINet, Faster R-CNN, and the improved ResNet-18 model (P<0.05). Overall, the optimized ResNet50 model exhibited exceptional performance in the early diagnosis of periodontal disease, highlighting its potential clinical application value.