RELA_Net: Upper Airway CBCT Image Segmentation Model Based on Receptive Field Expansion and Large-Kernel Attention

The structure of the upper airway is variable and complex due to its environmental and physiological factors. Currently, doctors mainly rely on manual outlining and segmentation from images. This method is time-consuming and relies heavily on the doctor’s experience. To solve this problem, we propose a fully automatic segmentation model for upper airway Cone Beam Computed Tomography (CBCT) images based on U-Net. The receptive field expansion module (RFEM) is used to replace the last three convolutional blocks of the encoder in the original U-Net model to improve the feature information extraction capability. And a large kernel attention module (LKA) is added to the skip connection part to dynamically adjust the receptive field of the feature extraction backbone, to alleviate the feature loss and redundancy of the skip connection. The dataset used in this paper is one created by us and the clinicians themselves, totaling 1345 CBCT images. Which were taken from 53 patients with airway obstruction. The imaging experts guided and delineated the label images. Experimental results show that the IoU and Dice score of the upper airway segmentation predicted by the RELA_Net network model in this article on the test sets are 94.39% and 97.10% respectively. Based on the prediction maps of the test set images, the segmentation model proposed in this article demonstrates an improvement in comparison to U-Net and other models, particularly in reducing over- and under-segmentation in the upper airway. This contributes to improving the diagnostic accuracy for patients with airway obstruction, thereby enhancing patient care and treatment planning.