Application of artificial intelligence in airway evaluation of cone beam computed tomography scans

Introduction: Multiple studies have supported the idea that new AI-based technologies can provide airway analysis comparable to that of human practitioners. Tested products include ‘Mimics’ (Materialise N.V., Leuven, Belgium) and ‘Diagnocat’ (Diagnocat, USA). According to the most comprehensive systematic review performed of these studies, “in order to have strong evidence on how accurately CBCT airway analysis is performed with the use of artificial intelligence, more randomized prospective studies should be performed. Those studies should limit all kind of bias to provide high quality evidence” (Tsolakis et al., 2022). Current literature has focused on testing these new technologies in very rigid conditions with the largest of these studies excluding datasets with artifact completely (Orhan et al., 2022). This is important because other reviews of these technologies have found that “overfitting is a common issue in the whole field of AI. This means the model performs excessively well in the training datasets but shows unsatisfactory performance in the testing dataset. Factors like data insufficiency, low data heterogeneity and excessive variables could all lead to overfitting” (Liu et al., 2023).Materials and Methods: 100 anonymized cone beam computed tomography datasets obtained from the records of the University of Iowa were oriented and analyzed by two calibrated oral and maxillofacial radiology residents using InVivo software (InVivo6, Anatomage, San Jose, CA). Measurements made were total airway volume and minimum surface area of the airway. These measurements were then compared to those produced by uploading the same datasets to Diagnocat AI software (Diagnocat, San Francisco, CA).