A study on a screening and diagnostic model for adenoid faces based on an improved convolutional neural network

To develop an automated, efficient model for screening and diagnosing adenoid faces by integrating multiple advanced Convolutional Neural Networks (CNNs), thereby providing a scientific basis for clinical diagnostics and large-scale epidemiological screening. This case-control study recruited 325 patients with adenoid faces and 445 controls with normal facial morphology from the Department of Pediatric Dentistry at the Stomatology Hospital of Yan’an University Affiliated Hospital (January 2024 - December 2024). Frontal and lateral photographs were collected from all participants (aged 11–14), with no significant demographic differences in age or gender between the groups (P > 0.05). We extracted facial features from the images to build seven distinct models: ViT, CrossViT, SimpleViT, DeepViT, ResNet34, VGG19_BN, and MedMamba. Model performance was rigorously evaluated on a test set using accuracy, precision, sensitivity, F1-score, and specificity, complemented by Precision-Recall (PR) and Receiver Operating Characteristic (ROC) curve analyses. The MedMamba model outperformed all other models, achieving an accuracy of 0.933, precision of 0.909, sensitivity of 0.975, F1-score of 0.941, and specificity of 0.882. ROC and PR curve analyses confirmed the robust diagnostic power of all seven models. The MedMamba and ResNet34 models exhibited the highest classification performance, with AUC values of 0.957 and 0.955, respectively. These findings position the MedMamba model as a highly effective preliminary screening tool for large-scale epidemiological surveys, capable of rapidly identifying suspected cases and optimizing resource allocation for subsequent professional assessment. The proposed adenoid faces screening model, built upon an integration of advanced CNNs, exhibits high accuracy and stability. It offers a valuable,objective decision-support tool for clinicians and establishes a solid foundation for future large-scale epidemiological research.