Dual-Branch Reconstruction Pre-Training with Multi-Layer Supervision and Regional Dynamic Graph Convolution for 2D Tooth Segmentation

Segmentation of teeth in 2D oral CT scans is crucial for diagnosing dental conditions. However, numerous existing models for 2D tooth segmentation fail to adequately account for the significance of tooth edge features in their design. Furthermore, these methods have overlooked the fact that oral images can be divided into regions, which affects the overall performance. To address these challenges, we introduce a new framework called Dual-Branch Reconstruction Pre-training with Multi-Layer Supervision and Regional Dynamic Graph Convolution (DBRP-MLS-RDGC) framework for 2D tooth segmentation. Specifically, the DBRP employs two branch systems, i.e., global image reconstruction and tooth edge reconstruction, as proxy tasks to pretrain the segmentation model, enabling it to capture both global image features and tooth edge features. Then, with the divided feature maps divided into the tooth, maxillary and mandibular regions, dynamic image convolution is applied to amplify the feature distinctions between the tooth region with surrounding tissues in the feature map. We carried out experiments using two publicly available 2D oral CT image datasets, demonstrating that the proposed approach exceeds current state-of-the-art models in performance.