Collaborative Tooth Motion Diffusion Model in Digital Orthodontics

Authors: Yeying Fan, Guangshun Wei, Chen Wang, Shaojie Zhuang, Wenping Wang, Yuanfeng Zhou

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive qualitative and quantitative experiments demonstrate the superiority of our framework in the application of orthodontics compared with state-of-the-art methods.
Researcher Affiliation Academia 1School of Software, Shandong University, China 2Department of Computer Science, The University of Hong Kong, China 3Texas A&M University, USA
Pseudocode No The paper does not include pseudocode or clearly labeled algorithm blocks.
Open Source Code No The paper does not provide an explicit statement about releasing the source code or a direct link to a code repository for the methodology.
Open Datasets No Our dataset consists of 1050 dental cases. Each dental case includes a group of models before and after alignment, the corresponding whole tooth motion process, and the tooth frames. These data are collected from hospitals and labeled by professional dentists.
Dataset Splits Yes We randomly divided the data into three parts for network training: 787 for training, 105 for validation, and 158 for testing.
Hardware Specification Yes All of them are trained on a single NVIDIA RTX 4090 GPU.
Software Dependencies No The paper mentions general software components like 'neural network learning' and 'PyTorch' implicitly but does not provide specific version numbers for any software dependencies.
Experiment Setup Yes Our model is trained with N = 100 noising steps and a cosine noise schedule. All of them are trained on a single NVIDIA RTX 4090 GPU. We use zero padding to deal with the missing teeth and downsample M = 400 points for each tooth point cloud from the intraoral scan model by farthest point sampling. The dimension of the tooth latent shape code Z is z = 16. The denoising network is a GRU with four hidden layers and a latent dimension 256. The empirical parameter δ is 0.1. We weigh our loss terms by λ1 = 1, λ2 = 1, λ3 = 10, λ4 = 1, λ5 = 0.1, and λ6 = 0.1. The framework is trained in two stages. The encoding part was trained first, and then its output was used as input for training the motion generation part.