Identifying Upper Airway and Evaluating Adenoid in Lateral Cephalometric Radiographs of Pediatric Patients Using Image-based Deep Learning Technique

Abstract

PurposeAdenotonsillar hypertrophy is considered as one of the primary causes of pediatric obstructive sleep apnea. In clinical practice, pediatric patients with potential obstructive apnea should undergo assessment of upper airway obstruction and adenotonsillar size. As lack of well-trained physicians in China, large numbers of such patients could not be accurately evaluated in time. We attempted to find a better way to assess upper airway obstruction. MethodsWe developed a computational method which may allow clinicians to trace upper airway obstruction. We utilized a Detectron2 Mask R-CNN architecture that was pretrained on ImageNet-1k dataset. We then trained it on our COCO dataset using transfer learning for segmentations of upper airway and related key structures. With the instance segmentations, we utilized a tracing algorithm developed by our own to sketch the contours of key segments. At the end, our system would use the traced landmarks to calculate the targeted clinical data. ResultsWe validated the effective-ness of the algorithm in two steps. First, we used a validation set (COCO dataset) to evaluate the segmentation performance of our system. Our system achieved a mean segmentation AP of 61.30, with airway segmentation AP of 55.64 and cranial base segmentation AP of 66.96. In addition, the AP50 was 99.49 and AP75 was 73.62. Second,we analyzed the data from our system and experts using a single rater, absolute-agreement, 2-way mixed-effects model, and got ICC value of 0.73 with 95% confident interval = 0.63-0.81. ConclusionsIn this study, we created a deep-learning based system to help clinicians evaluate upper airway in lateral cephalometric radiographs which we believe could improve clinical practice. With the evolving of technology, our system would become more integrated into medical care of OSA, freeing the clinical practitioners from repetitive tasks and enabling them to concentrate on improved patient care.