Multimodal Intelligent Diagnostic Model For Pediatric Sagittal Skeletal Patterns

To develop an intelligent diagnostic system for pediatric sagittal skeletal patterns based on multimodal imaging features, achieving synergistic optimization of radiation-free screening and precise diagnosis. A retrospective study included 1,613 pediatric patients from Guangxi Medical University School of Stomatology (January 2018–December 2022). Lateral cephalograms and profile photos were collected and labeled as skeletal Class I, II, or III by three orthodontic experts. The performance of four deep learning models (ResNet-101, DenseNet-121, Swin-T, and ConvNeXt-T) was compared to identify the optimal architecture. Separate unimodal diagnostic models were developed for lateral cephalograms and profile photos, followed by a multimodal feature alignment model. Five-fold cross-validation was used to assess model performance, with clinical validation on 300 independent cases, including 60 physician-AI comparative tests. Cross-validation showed that the diagnostic accuracy of the lateral cephalogram model reached 90.33%, the profile photo model 83.39%, and the multimodal model 88.51%, outperforming unimodal models by 5.12%. Clinical validation revealed that senior, mid-level, and junior physicians achieved 69.59%, 70.17%, and 76.60% accuracy with lateral cephalograms, while AI reached 89.95% (P<0.001). With profile photos, physicians achieved 62.33%–68.91% accuracy, whereas AI achieved 87.39% . This study established the first intelligent classification model integrating X-ray and radiation-free facial images for pediatric sagittal skeletal diagnosis, enhancing early screening and prevention of malocclusion by improving the timeliness and accuracy of skeletal anomaly recognition.