Machine learning for the automatic detection of adenoid hypertrophy based on lateral cephalograms

Objective To develop and validate a fully automatic artificial intelligence (AI)-driven system for children with an adenoid hypertrophy (AH) screening based on lateral cephalograms (LCs). Methods A total of 679 LCs obtained from 12-year-old children, including 582 Chinese and 97 Caucasian children, were randomly assigned to a training set (n = 459) to train AI models, a validation set (n = 100) to determine optimal parameter values and a test set (n = 120) to evaluate the AI model’s performance. A five-fold cross-validation was employed to evaluate accuracy. A category-based relation consistency mean teacher network (CRC-MT) model was trained to automatically screen AH. Accuracy, recall, precision and F1 score were used to evaluate the performance of the AI model for AH screening. The AH screening performances of the AI-driven models and manual processing were compared. The accuracies of AI-driven and manual AH detection with blocked regions were also analysed and compared. Results The accuracy and recall of AH screening in full LCs were 80.5%–85.6% and 69.5%–79.5%, respectively; the precision and F1 scores were 67.7%–75.7% and 66.0%–85.9%, respectively. The chi-square test revealed that the AI-driven model achieved higher accuracy in terms of the accuracy, recall, precision and F1 score of AH screening compared with human results. The accuracy of AH screening in the blocked adenoid, airway and jaw was 0.820, 0.883 and 0.892, respectively. The performance of manual AH screening was significantly low in the blocked adenoid, upper airway and jaw region images. Conclusions A fully automatic AI-driven system was developed and validated for the screening of AH in children based on LCs.