Development of screening patients with obstructive sleep apnea with deep learning in the lateral cephalogram: Overcoming modality differences with knowledge distillation

Undetected obstructive sleep apnea (OSA) can lead to consequences of severe systematic disease. The lateral cephalogram in orthodontics is a valuable screening tool. We hypothesized that a deep learning-based classifier might be able to differentiate sleep apnea as anatomical features that humans do not recognize in lateral cephalogram. Moreover, since the imaging devices used by each hospital could be different, various modalities in radiography need to be overcome in real clinical practice. Therefore, we proposed a knowledge distillation deep learning model to classify patients into OSA and non-OSA groups using the lateral cephalogram and to overcome modality differences simultaneously. Lateral cephalograms of 500 OSA patients and 500 non-OSA patients from two different devices were included. ResNet-50 and ResNet-50 with a feature-based knowledge distillation model were trained and their suitability for classification and modality normalization were compared. Through knowledge distillation, it was confirmed through ROC analysis and Grad-CAM that our model exhibits high performance without being deceived by features caused by modality differences. By checking the probability values predicting OSA, an improvement in overcoming the modality differences in lateral cephalogram was observed, which could be applied in the actual clinical situation.