ToothFussion: Generate a completed tooth from crown information
Published:
Project Objective:
To generate complete CBCT tooth structures from partial IOS crown information and textual descriptions, thereby reducing dental patients’ exposure to CT radiation and lowering the risk of radiation-induced malignancies.
Main Output:
○ Trained an IOS segmentation model on over 3,000 patient OBJ files (accuracy >97%) and successfully applied it to 145 new patient cases to obtain high-quality segmented data; ○ Performed full-mouth registration of CBCT/IOS STL files using CloudCompare and the ICP algorithm, computed and applied transformation matrices to individual teeth, and constructed CBCT/IOS single-tooth registration pairs; ○ Fine-tuned a VQ-VAE 3D latent compression model on 2,000+ single-tooth datasets to better adapt it for tooth morphology modeling; ○ In Progress: Training a multimodal diffusion model (3D data + text conditions) with GPT and 3D U-Net as the primary architectures, aiming to generate complete tooth shapes from partial crown geometries.
项目截图
架构图
实验结果
