Integrating Spatial Contrast Attention and Triple-branch Tunable Atrous Convolution for Panoramic Dental X-ray Segmentation

Accurate tooth segmentation in panoramic dental X-ray images is crucial for early diagnosis and orthodontic planning, yet overlapping teeth, metal artifacts, and image noise make this task exceptionally challenging. To address these issues, we propose the Mamba Integration Network (MambaTriNet), which integrates three key modules: Spatial Contrast Attention (SCA), employing both max-pooling and average-pooling to generate spatial attention maps that enhance salient features while suppressing noise; Triple-branch Tunable Atrous convolution (TTA), which uses three parallel dilated-convolution branches with adaptive dilation rates and enhanced local attention to capture multi-scale contextual information and long-range dependencies; and Multi-Layer Convolutional Fusion (MLCF), which combines SCA with a two-dimensional state-space model (SS2D) to strengthen edge detection and improve robustness against metal artifacts. Comprehensive experiments on the MICCAI tooth segmentation dataset show that MambaTriNet outperforms U-Net, TransUNet, Mamba-UNet, and VMUNet across multiple metrics—achieving a Dice score of 93.10%, an IOU of 91.86%, and a 95% Hausdorff distance of 7.324—thereby demonstrating its effectiveness and robustness in preserving fine details and precisely segmenting tooth boundaries.