<p>Accurate segmentation of multiple oral and maxillofacial structures in Cone Beam Computed Tomography (CBCT) images has become increasingly important for clinical diagnosis and treatment planning. Despite substantial advances in existing studies, accurately segmenting boundary regions where structures exhibit highly similar characteristics remains challenging. In addition, the imbalance in class distributions severely degrades the segmentation precision in low-frequency classes. To address these challenges, this paper proposes a multi-class segmentation algorithm for oral and maxillofacial structures in CBCT images that integrates a two-level attention mechanism and an optimized loss. Firstly, a Multi-scale Perception Reverse Attention module was designed. Leveraging a reverse attention mechanism, it uses the multi-scale fused features as auxiliary information to sharpen the model’s focus on complementary regions and fine details, thereby significantly enhancing feature representation. Secondly, a Laplacian Edge-enhanced Gated Attention module is constructed, in which the edge features extracted by the 3D Laplacian pyramid algorithm are fused with the encoder output features and subsequently fed into the gated attention module. This design strengthens the model’s focus on critical edge information and further improves segmentation performance. Finally, a Class Frequency-guided Weighted Loss is proposed to mitigate class imbalance, promoting the model’s focus on less frequent classes while preserving accuracy for other classes. Experimental results from ToothFairy2 and the private dataset show that the proposed method outperforms most existing algorithms, achieving excellent segmentation performance.</p>