Deep Learning Models for Automatic Upper Airway Segmentation and Minimum Cross-Sectional Area Localisation in Two-Dimensional Images

Objective: To develop and validate convolutional neural network algorithms for automatic upper airway segmentation and minimum cross-sectional area (CSAmin) localisation in two-dimensional (2D) radiographic airway images.

Materials and methods: Two hundred and one 2D airway images acquired using cone-beam computed tomography (CBCT) scanning were randomly assigned to a test group (n = 161) to train artificial intelligence (AI) models and a validation group (n = 40) to evaluate the accuracy of AI processing. Four AI models, UNet18, UNet36, DeepLab50 and DeepLab101, were trained to automatically segment the upper airway 2D images in the test group. Precision, recall, Intersection over Union, the dice similarity coefficient and size difference were used to evaluate the performance of the AI-driven segmentation models. The CSAmin height in each image was manually determined using three-dimensional CBCT data. The nonlinear mathematical morphology technique was used to calculate the CSAmin level. Height errors were assessed to evaluate the CSAmin localisation accuracy in the validation group. The time consumed for airway segmentation and CSAmin localisation was compared between manual and AI processing methods.

Results: The precision of all four segmentation models exceeded 90.0%. No significant differences were found in the accuracy of any AI models. The consistency of CSAmin localisation in specific segments between manual and AI processing was 0.944. AI processing was much more efficient than manual processing in terms of airway segmentation and CSAmin localisation.

Conclusions: We successfully developed and validated a fully automatic AI-driven system for upper airway segmentation and CSAmin localisation using 2D radiographic airway images.