Multi-Path Restorative Pre-Training with Localized Adaptive Graph Convolution for 2D Tooth Segmentation

Segmenting teeth in 2D oral CT images is crucial for diagnosing dental conditions. However, many current models for 2D tooth segmentation depend solely on end-to-end segmentation loss during training, which can fail to capture the image's intrinsic features. Additionally, neighboring regions may interfere with tooth segmentation, leading to less-than-optimal results. To address these challenges, we introduce a novel framework named Multi-Path Restorative Pre-training with Localized Adaptive Graph Convolution (MPRP-LAGC) for 2D tooth segmentation. The MPRP utilizes two paths - one for image reconstruction and another for tooth edge reconstruction - as auxiliary tasks for pretraining the segmentation model. This enables the model to better capture both global image characteristics and the specific features of tooth edges. After dividing the feature maps to correspond to the tooth, maxillary, and mandibular regions, dynamic convolution is applied to emphasize the distinctions between the tooth area and surrounding tissues in the feature space. Experiments on one publicly available 2D oral CT dataset show that the proposed method surpasses current state-of-the-art models and accurately segments the boundaries of teeth, particularly where they meet the maxilla and mandible.