Comparison of Convolutional Neural Networks Architectures for Screening Mandibular-Plane-to-Hyoid Distance as a Risk Indicator of Obstructive Sleep Apnea on Lateral Cephalograms

Abstract

Background: Lateral cephalometric radiographs can aid in screening obstructive sleep apnea (OSA) through the mandibular-plane-to-hyoid (MP-H) distance. An automated tool based on this measure could support OSA screening. However, the accuracy of such tools depends on the performance of AI-assisted models, making it essential to evaluate and compare their effectiveness in detecting MP-H distance for reliable clinical application. Objective: To evaluate the performance of different convolutional neural network (CNN) architectures in classifying patients into short and long MP-H groups. Materials and methods: A total of 304 pre-orthodontic lateral cephalometric radiographs from adults (age 18-56 years) were classified into short MP-H (< 18 mm) and long MP-H (≥ 18 mm) groups. Four CNN architectures (DenseNet-121, ResNet-50, EfficientNet-B0, and MobileNetV3) were trained to classify short and long MP-H groups. To address class imbalance, weighted binary cross entropy loss functions (weights ranging from 1 to 5) were applied, assigning greater penalties to misclassification of the minority class. Results: In the scenario without application of weighted cross entropy loss, DenseNet-121 achieved the overall high screening performance, with sensitivity = 0.87, specificity = 0.95, precision = 0.79, F1-score = 0.82, accuracy = 0.93, AUROC = 0.91, and AUPRC = 0.83. MobileNetV3 consistently demonstrated the lowest performance. Weighted loss functions provided inconsistent benefits across architectures. DenseNet-121 showed consistent performance among all weights. Conclusion: DenseNet-121 shows potential of screening long MP-H distance from lateral cephalograms. Weighted loss functions may provide improvements, but model selection remains the more critical factor.