TranSDFNet: Transformer-Based Truncated Signed Distance Fields for the Shape Design of Removable Partial Denture Clasps.

Xinze Shen,Changdong Zhang,Xiuyi Jia,Dawei Li,Tingting Liu,Sukun Tian,Wei,Yuchun Sun,Wenhe Liao

DOI: https://doi.org/10.1109/jbhi.2023.3295387

IF: 7.7

2023-01-01

IEEE Journal of Biomedical and Health Informatics

Abstract:The ever-growing aging population has led to an increasing need for removable partial dentures (RPDs) since they are typically the least expensive treatment options for partial edentulism. However, the digital design of RPDs remains challenging for dental technicians due to the variety of partially edentulous scenarios and complex combinations of denture components. To accelerate the design of RPDs, we propose a U-shape network incorporated with Transformer blocks to automatically generate RPD clasps, one of the most frequently used RPD components. Unlike existing dental restoration design algorithms, we introduce the voxel-based truncated signed distance field (TSDF) as an intermediate representation, which reduces the sensitivity of the network to resolution and contributes to more smooth reconstruction. Besides, a selective insertion scheme is proposed for solving the memory issue caused by Transformer blocks and enables the algorithm to work well in scenarios with insufficient data. We further design two weighted loss functions to filter out the noisy signals generated from the zero-gradient areas in TSDF. Ablation and comparison studies demonstrate that our algorithm outperforms state-of-the-art reconstruction methods by a large margin and can serve as an intelligent auxiliary in denture design.