This study aimed to develop and validate a multi-view convolutional neural network (CNN) for detecting craniosynostosis in pediatric skull radiographs. Radiographs from 996 children under 36 months of age at Seoul National University Children’s Hosp...
This study aimed to develop and validate a multi-view convolutional neural network (CNN) for detecting craniosynostosis in pediatric skull radiographs. Radiographs from 996 children under 36 months of age at Seoul National University Children’s Hospital were retrospectively collected and divided into training/validation (n = 846) and internal test (n = 150) cohorts, with external validation at Pusan National University Hospital (n = 104) and a simulated retrospective cohort (n = 97). The model incorporated anteroposterior or posteroanterior, Towne’s, and lateral views, reflecting the integrative diagnostic approach of radiologists. Single-view CNN encoders were pretrained and fine-tuned in a multi-view architecture using focal loss and test-time augmentation. Model performance was evaluated by receiver operating characteristic (ROC) analysis and standard classification metrics at both the best-performance and 100%-sensitivity thresholds. The proposed model achieved AUCs of 0.999, 0.987, and 0.994 in the internal, external, and retrospective cohorts, respectively. At the best threshold, sensitivity and specificity were 98.1% and 99.0% (internal), 100% and 84.4% (external), and 83.3% and 98.8% (retrospective). In the retrospective cohort set, the model could have reduced unnecessary CT examinations by 75% in the screening setting and 100% in the stand-alone setting. Grad-CAM visualization revealed attention patterns concordant with radiologic reasoning, while misclassifications were mainly attributed to attention dropout or misalignment. This multi-view deep-learning model demonstrates excellent diagnostic performance and interpretability, suggesting its potential as a prescreening or triage tool to reduce radiation exposure and improve diagnostic efficiency in the evaluation of craniosynostosis.