Application of deep learning in the radiographic image analysis of upper airway and adenoid

Abstract

Obstructive sleep apnoea (OSA) is a sleep disorder characterised by recurrent episodes of partial or complete upper airway obstruction during sleep. Recurrent interruptions in breathing during sleep lead to disrupted sleep patterns and inadequate oxygen supply. In undiagnosed and untreated adults, OSA is associated with hypertension, coronary artery disease, cardiac arrhythmias and depression. Persistent respiratory obstruction in children is associated with various craniofacial deformities, such as a long face, constricted maxillary arch, narrow nasal base and backward rotation of the mandible. Currently, polysomnography is the standard method used for diagnosing OSA; however, it is complex, time-consuming and expensive. Traditional radiographic examinations are basic tools for preliminary upper airway assessment, and numerous studies have indicated a correlation between airway morphology and adenoid hypertrophy (AH) and respiratory obstruction. Analysis of upper airway and adenoid structures using radiographic images can help healthcare professionals, particularly dentists, to identify potential problems in airway dimensions and certain anatomical features or abnormalities that could be associated with respiratory diseases. In recent years, artificial intelligence (AI) techniques have gained significant traction across various fields. We reviewed the clinical applications of dental image-based machine learning methods in orthodontics during the past decade, as detailed in Chapter I. We found that the application of AI techniques in upper airway and adenoid assessment has not been thoroughly explored. Therefore, we explored the possibilities of applying AI techniques to automatic upper airway and adenoid assessment to achieve more accurate and efficient evaluations. In this study, we developed and validated fully automatic AI-driven systems for upper airway segmentation and minimum cross-sectional area (CSAmin) localisation in two-dimensional (2D) radiographic airway images (Chapter III). Furthermore, we used cone-beam computed tomography images as training data to develop AI models for upper airway segmentation and CSAmin localisation in three dimensions (Chapter IV). In both 2D and 3D studies, the AI models achieved a segmentation accuracy exceeding 90.0%, and the height difference error between AI processing and human annotation was within 3 mm. In terms of efficiency, AI processing performed much better than manual processing in both tasks. In addition to upper airway assessment, we established an AI model for AH diagnosis in children (Chapter V). The category-based relation consistency mean teacher network (CRC-MT) diagnostic model was established for AH diagnosis based on 679 lateral cephalograms obtained from 12-year-old Chinese and Caucasian children. In comparison with manual methods, AI methods produced results with a significantly higher accuracy across evaluation metrics. Our blocked region test results showed that AI methods can detect features and characteristics from images that may not be discernible to human eyes. In conclusion, our study contributes to the advancement of AI applications in upper airway and adenoid imaging processing, with potential implications for broader applications in orthodontics. Regarding the applications of AI, upper airway and adenoid imaging assessments are basic and initial steps in the development of AI systems for further pathology detection and disease diagnosis. From a clinical perspective, our study introduces an innovative AI-driven method that will help improve clinicians’ efficiency.