Digital dentistry has been upgraded traditional dentistry by realizing the complete process linking preliminary image processing tasks to orthodontic diagnosis and treatment planning.Automatic tooth labeling from dent...Digital dentistry has been upgraded traditional dentistry by realizing the complete process linking preliminary image processing tasks to orthodontic diagnosis and treatment planning.Automatic tooth labeling from dental images involves identifying a specific tooth class using deep learning algorithms.Existing methods used for tooth labeling often use incomplete tooth models,with only crown information obtained by oral scanners or X-ray radiographs.Even with cone beam computed tomography(CBCT)images,the problem of misclassification with a change in tooth orientation still appears.Consequently,they could not perform well for the varying test samples with similar teeth or those with missing teeth.We propose a tooth labeling method based on complete tooth models using a few-shot classification based on a meta-learning approach.To resolve the problem of intra-class similarities,we design an attention mechanism based on matrix decomposition inspired by HamNet.We validate the proposed method on the CBCT dataset with augmented tooth models.The results demonstrate that our approach performs better than cutting edge methods like PointNet,PointCNN,PointNet++,dynamic graph convolutional neural network(DGCNN),hierarchical convolutional neural network(HCNN),and TSegNet in terms of accuracy with a significant improvement of 11.46%,7.8%,6.16%,4.48%,5.06%and 5.08%,respectively.Therefore,with validated data applicability,automated tooth labeling will be applied in digital dentistry with much excellence and feasibility.展开更多
基金supported by the National Natural Science Foundation of China(Nos.U2013205 and 62073309)the Chinese Academy of Sciences Youth Innovation Promotion Association Excellent Member Program,China(No.Y201968).
文摘Digital dentistry has been upgraded traditional dentistry by realizing the complete process linking preliminary image processing tasks to orthodontic diagnosis and treatment planning.Automatic tooth labeling from dental images involves identifying a specific tooth class using deep learning algorithms.Existing methods used for tooth labeling often use incomplete tooth models,with only crown information obtained by oral scanners or X-ray radiographs.Even with cone beam computed tomography(CBCT)images,the problem of misclassification with a change in tooth orientation still appears.Consequently,they could not perform well for the varying test samples with similar teeth or those with missing teeth.We propose a tooth labeling method based on complete tooth models using a few-shot classification based on a meta-learning approach.To resolve the problem of intra-class similarities,we design an attention mechanism based on matrix decomposition inspired by HamNet.We validate the proposed method on the CBCT dataset with augmented tooth models.The results demonstrate that our approach performs better than cutting edge methods like PointNet,PointCNN,PointNet++,dynamic graph convolutional neural network(DGCNN),hierarchical convolutional neural network(HCNN),and TSegNet in terms of accuracy with a significant improvement of 11.46%,7.8%,6.16%,4.48%,5.06%and 5.08%,respectively.Therefore,with validated data applicability,automated tooth labeling will be applied in digital dentistry with much excellence and feasibility.
