Automated segmentation of maxillary sinus from computed tomography images: A comparative analysis of U-net and its variants with pre-trained encoders

The maxillary sinus holds an important position in the respiratory system, and its accurate segmentation from computed tomography (CT) images remains challenging. Traditional manual and semi-automatic methods are time-consuming and inconsistent, making them less suitable for clinical applications. Automated segmentation has become possible thanks to the advancements in deep learning, but challenges still exist due to the complex sinus anatomy and inter-individual variations. This study presents the development and evaluation of four different deep learning models for maxillary sinus segmentation: a plain U-Net model, and three U-Net variations that utilize VGG16, ResNet50, and Inception-ResNet v2 as encoders. The main objectives are to evaluate these models for accurate segmentation, to further segment the air section and mucosal inflammation within the MS, to calculate the maxillary sinus volume and opacification ratios, and to demonstrate the potential of deep learning models in maxillofacial radiology. The Inception-ResNet v2 - UNet model demonstrated the best performance with a Dice score of 0.96671 and an Intersection over Union (IoU) score of 0.94004, while all models, including the plain U-Net, have shown noteworthy successes.