Résumé: Accurate identification of the maxillary sinus and maxillary sinus septum on cone beam computed tomography (CBCT) is essential for appropriate surgical methods. However, the large-scale gap and the prior knowledge between the targets make the general medical image segmentation model unable to precisely segment. This paper proposes a novel axial-attention DBA-UNet for automatic maxillary sinus and maxillary sinus septum segmentation. The proposed DBA-UNet consists of a two-stage network based on the U-Net backbone. Firstly, to improve the segmentation precision at the target’s boundary, a boundary Cross-fusion network is constructed to segment the maxillary sinus and its boundary. Secondly, in the second stage, an axial-attention Transformer fusion strategy is proposed to utilize the prior knowledge and long-range dependencies between the targets. By fusing features of the maxillary sinus boundary and the original CBCT images, the boundary attention network can improve the segmentation performance of the maxillary sinus septum. The experimental results demonstrate that the proposed DBA-UNet outperforms the network without boundary Cross-fusion and boundary attention and can improve the average Dice similarity coefficient and Hausdorff distance effectively, which indicates that the proposed method is effective and of great clinical significance.

Langue: Anglais