Abstract
This study proposed and investigated the performance of a deep learning based three-dimensional (3D) convolutional neural network (CNN) model for automatic segmentation of the pharyngeal airway space (PAS).
A dataset of 103 computed tomography (CT) and cone-beam CT (CBCT) scans was acquired from an orthognathic surgery patients database. The acquisition devices consisted of 1 CT (128-slice multi-slice spiral CT, Siemens Somatom Definition Flash, Siemens AG, Erlangen, Germany) and 2 CBCT devices (Promax 3D Max, Planmeca, Helsinki, Finland and Newtom VGi evo, Cefla, Imola, Italy) with different scanning parameters. A 3D CNN-based model (3D U-Net) was built for automatic segmentation of the PAS. The complete CT/CBCT dataset was split into three sets, training set (n = 48) for training the model based on the ground-truth observer-based manual segmentation, test set (n = 25) for getting the final performance of the model and validation set (n = 30) for evaluating the model's performance versus observer-based segmentation.
The CNN model was able to identify the segmented region with optimal precision (0.97±0.01) and recall (0.96±0.03). The maximal difference between the automatic segmentation and ground truth based on 95% hausdorff distance score was 0.98±0.74mm. The dice score of 0.97±0.02 confirmed the high similarity of the segmented region to the ground truth. The Intersection over union (IoU) metric was also found to be high (0.93±0.03). Based on the acquisition devices, Newtom VGi evo CBCT showed improved performance compared to the Promax 3D Max and CT device.
The proposed 3D U-Net model offered an accurate and time-efficient method for the segmentation of PAS from CT/CBCT images.
The proposed method can allow clinicians to accurately and efficiently diagnose, plan treatment and follow-up patients with dento-skeletal deformities and obstructive sleep apnea which might influence the upper airway space, thereby further improving patient care.