RTNet: The Network Model for Refining Tooth X-ray Image Segmentation Boundaries

In regions with a high incidence of dental diseases, more intelligent treatment methods can enhance treatment efficiency and accuracy. Automatic and accurate tooth segmentation is a crucial component of dental intelligence. However, due to the weak contrast in dental X-ray images, it results in difficulty in segmentation using deep learning methods, leading to a low-quality Mask obtained. In this paper, we propose RTNet to improve the quality of tooth X-ray image segmentation. The model is an overall U-shaped structure, and on the skip connections of the external U-shaped structure, we introduce Triple Spelling Attention, which reduces the negative impact of the encoder on the decoder and retains highly similar information. On the internal U-shaped structure, we use offset maps to calibrate the up-sampled features, enhancing the segmentation performance of the boundary region. For refining the mask, we compute feature information for error-prone points and employ a refinement network for learning, improving mask quality. Experimental results demonstrate that our proposed model exhibits strong competitiveness compared to other segmentation models.