Exécution de notre outil de synthèse sur un nouvel article
Vérifiez l’article ci-dessous, puis lancez sa synthèse.
DilatedToothSegNet++: topology-aware graph neural network for robust 3D dental mesh segmentation
Dong Wang (Foshan University), Fenquan Ye (Foshan University), Rongjian Xu (Foshan University), Dongmei Lin (Foshan University)
Résumé : Segmenting individual teeth from 3D intraoral scans is a critical step in digital dentistry workflows. While deep learning has advanced 3D surface segmentation considerably, dental data present persistent difficulties—anatomical variations between patients, missing teeth, and positional irregularities create challenges. We introduce DilatedToothSegNet++, a graph-neural architecture designed for precise dental mesh segmentation. Three innovations distinguish this model: geometric-curvature descriptors enhance shape discrimination; topology-sensitive neighbor selection reduces cross-tooth interference; an adaptive multi-scale refinement module accommodates structural variability and corrects scan artifacts. Performance on the Teeth3DS benchmark reaches 96.24% overall accuracy and 88.15% mIoU (mean Intersection‑over‑Union). Experimental results demonstrate that the model remains reliable in complex clinical cases.