Development and validation of a vision transformer model for radiographic assessment of pediatric humeroradial joint alignment.

BACKGROUND: Whether the radiographic assessments of pediatric anatomical (A-HRJA) and nonanatomical humeroradial joint alignment (NA-HRJA) among pediatric orthopedic surgeons (POS) relies on clinical experience remains unknown. No vision transformer (ViT) model for automated and accurate radiographic pediatric HRJAs categorization has been developed and validated. This study aims to evaluate the radiographic assessment outcomes of pediatric A-HRJAs and NA-HRJAs among POS with varying levels of experience, and to develop and validate a model for automated and accurate categorization of pediatric HRJAs on radiographs. MATERIALS AND METHODS: A total of 294 pediatric radiographs were independently recorded by four junior-level, four mid-level, and three senior-level POS. Then, the performance of the POS with various levels of experience was evaluated. A total of 5118 pediatric radiographs (2932 A-HRJAs and 2186 NA-HRJAs) were prepared to construct the vision transfer (ViT) model. The ViT model's clinical efficacy was evaluated using an external testing dataset (84 A-HRJAs and 132 NA-HRJAs). A total of 11 POS independently recorded the HRJA types of the external testing dataset twice with a 1-month interval: first without ViT model assistance, and second with assistance. The rating performances of the POS without and with ViT model assistance were compared. RESULTS: The accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were similar among junior-level, mid-level, and senior-level POS (P > 0.050). The ViT model's area under the receiver operating characteristic curve was 0.944. The ViT model's rating performance was significantly superior to that of POS in the first round regarding accuracy (P = 0.001), sensitivity (P = 0.001), and NPV (P < 0.001). However, no significant differences were found between the ViT model and POS in the second round. The accuracy (P = 0.008), PPV (P = 0.043), and NPV (P = 0.030) improved significantly in the second round compared with the first, as demonstrated by the following values: 86.9-92.6%, 88.3-93.6%, and 85.8-93.2%, respectively. The kappa value reflecting the reliability among POS improved by 24.8% in the second round. CONCLUSIONS: The accuracy of manually classifying radiographic pediatric HRJA types did not depend on clinical experience. The ViT model can accurately categorize radiographic pediatric HRJA types and significantly improve the POS's rating performance when categorizing A-HRJAs and NA-HRJAs. LEVEL OF EVIDENCE: III.