Automatic Segmentation of Maxillary Sinus with U-Net Model with Pre-trained Encoder

Abstract

An accurate segmentation of the maxillary sinus (MS) is crucial for the preoperative planning of MS-related surgeries and for preventing postoperative complications. Manual segmentation is challenging, time-consuming, and highly dependent on the practitioner's experience. Therefore, it is not applicable for clinical practice, and accurate, efficient automatic segmentation of MS is required. Convolutional neural networks (CNNs) have recently become the most preferred method for automatic medical image segmentation. In this study, an automatic MS segmentation model based on a convolutional neural network model, U-Net, is proposed. Instead of using the original U-Net encoder, the VGG16 network pre-trained with the ImageNet dataset, apart from the fully connected layers, was used as the encoder of the U-Net architecture to improve the segmentation accuracy. Furthermore, during the training period, models were also trained with focal dice loss (FDL), an equally weighted combination of dice loss (DL) and focal loss, to overcome the imbalanced dataset. The segmentation model based on U-Net with a VGG16 encoder trained with FDL obtained the best results with a dice similarity coefficient (DSC) of 0.93253 and an Intersection over Union (IoU) of 0.88775 on the test dataset.