파노라마 X-ray 영상에서의 상악동 영역 세그멘테이션을 위한 어텐션 기반의 U-Net 트랜스포머 (Attention-based U-Net Transformer for Segmentation of Maxillary Sinus Regions in Panoramic X-ray Images)

An important process in dental implant surgery is localizing maxillary sinus regions to ensure proper implant placement. Panoramic radiography images have been widely utilized to analyze the anatomy and position of the maxillary sinus. Previous studies have employed manual and semi-automatic segmentation methods to localize regions of interest in radiography images. However, manual analysis of maxillary sinus regions requires expert knowledge and is a time-consuming task. To address these challenges, we propose a deep learning-based segmentation algorithm for analyzing the shape and position of maxillary sinus regions. The proposed model employs a U-shaped network with transformers as encoders and introduces an attention module to enhance the feature maps obtained from the encoder. We compute spatial saliency using depth-wise separable convolution and integrate it with a channel attention module, and it is referred to as depthwise separable attention module (DSAM). It reduces the number of parameters and improves performance compared to conventional convolution-based spatial attention. We evaluated our model on a real-world dental panoramic radiography dataset, and the results demonstrate that the proposed method improves maxillary sinus segmentation performance across all evaluation metrics.

An important process in dental implant surgery is localizing maxillary sinus regions to ensure proper implant placement. Panoramic radiography images have been widely utilized to analyze the anatomy and position of the maxillary sinus. Previous studies have employed manual and semi-automatic segmentation methods to localize regions of interest in radiography images. However, manual analysis of maxillary sinus regions requires expert knowledge and is a time-consuming task. To address these challenges, we propose a deep learning-based segmentation algorithm for analyzing the shape and position of maxillary sinus regions. The proposed model employs a U-shaped network with transformers as encoders and introduces an attention module to enhance the feature maps obtained from the encoder. We compute spatial saliency using depth-wise separable convolution and integrate it with a channel attention module, and it is referred to as depthwise separable attention module (DSAM). It reduces the number of parameters and improves performance compared to conventional convolution-based spatial attention. We evaluated our model on a real-world dental panoramic radiography dataset, and the results demonstrate that the proposed method improves maxillary sinus segmentation performance across all evaluation metrics.