Dual-Embedding Guided Adaptive Hypergraph ODE Network for Early Diagnosis and Risk Assessment of Pediatric Impacted...

Impacted maxillary anterior teeth in children may cause dental arch deformities and abnormal craniofacial development, making early detection and intervention essential for normal oral development. Recent deep learning-based diagnostic methods have shown promise in dental image analysis. However, they often struggle to capture the complex spatial relationships between teeth and surrounding anatomical structures, limiting their ability to accurately assess expert-defined at-risk teeth. To address this challenge, we propose the Dual-embedding guIded Adaptive hypergraph ODE Model (DIADEM). DIADEM adopts a hierarchical encoder-decoder architecture that integrates an adaptive hypergraph ODE with dual-embedding prior constraints. Given panoramic radiographs, the model performs precise segmentation of teeth, detects existing impacted teeth, and identifies expert-defined at-risk teeth, while providing both pixel-level delineation and segmentation-derived tooth-level diagnostic evaluation. Specifically, the adaptive hypergraph ODE (HG-ODE) module models high-order topological relationships by dynamically adjusting diffusion depth and feature coupling strength. In addition, the dual-embedding mechanism incorporates polar-Cartesian spatial priors and edge-aware gradient constraints to enforce geometric consistency between adjacent dental structures. Experimental results on the PDX-Seg dataset demonstrate that DIADEM achieves excellent performance in impacted teeth diagnosis, demonstrating its screening-support potential in both identifying already impacted teeth and assessing expert-defined at-risk teeth, supporting timely intervention in pediatric patients.