Deep learning-based automatic segmentation of maxillary sinus and maxillary posterior teeth in Cone-beam computed tomography images

LI Chengye; YI Kexin; ZHANG Mingming; LIANG Yuhong

doi:10.15956/j.cnki.chin.j.conserv.dent.2025.08.002

Abstract：Objective To develop an improved U-Net deep learning model integrated with the convolutional block attention module(CBAM)for segmenting maxillary sinus(MS)and maxillary posterior teeth(MPT)structures in cone-beam computed tomography(CBCT)images,and to evaluate the segmentation accuracy of the U-Net CBAM model.Methods Using 35 CBCT datasets(containing 70 MS cases and 280 MPT)annotated by endodontic specialists as the ground truth,a U-Net CBAM model incorporating dual channel-spatial attention mechanisms was developed.The model was trained using three-fold cross-validation,and its segmentation accuracy was evaluated and compared with U-Net and EfficientNet models.Evaluation metrics included Dice coefficient(DSC),Jaccard similarity index(JSI),and volume consistency metrics.Results All three models achieved DSC and JSI values above 0.90 and 0.84 for MS segmentation,and above 0.83 and 0.76 for MPT segmentation,respectively.The U-Net CBAM model demonstrated the highest performance,with DSC and JSI values of 0.935 and 0.895 for MS segmentation,and 0.887 and 0.796 for MPT segmentation.In comparison,the baseline U-Net showed lower accuracy,yielding DSC values of 0.902(MS)and 0.830(MPT),and JSI values of 0.842(MS)and 0.709(MPT).Volume consistency analysis revealed a strong linear correlation(R2＞0.9)between U-Net CBAM automated segmentation and manual annotations.Conclusion Compared to U-Net and EfficientNet,the enhanced U-Net model with CBAM exhibited superior performance in segmenting MS and MPT structures in CBCT images,confirming the effectiveness of channel-spatial attention mechanisms in identifying complex anatomical structures.This study provides a valuable framework for multi-target segmentation in small-sample oral imaging applications.