Abstract:
The segmentation of individual 3D teeth models is a key technique in computer-aided orthodontic systems.Due to the complexity of interactive operation and the high degree of manual intervention,the efficiency and accuracy of traditional tooth segmentation methods are low.Therefore,we propose a novel approach based on multi-level 3D convolution neural networks for segmenting and recognizing tooth types.Firstly,the labeling preprocessing for dental models is carried out by the constructed octree sparse representation model based on Hash table.Secondly,to reduce the misclassification in highly similar tooth categories,the general features are extracted from all tooth categories using Level-1 network,and the specific features are extracted from highly similar tooth categories by using Level-2 network.Finally,the multi-level hierarchical segmentation network based on the deep convolution features is used to conduct segmentation of teeth-gingiva and inter-teeth,and the conditional random field model is used to optimize the boundary of the gingival margin and the inter-teeth fusion region.Experimental results show that the classification accuracy on the self-collected dental dataset is maintained above 0.858,and the accuracy of single tooth segmentation is 0.898.Compared with the similar segmentation methods,it is verified that the hierarchical feature learning method is robust and accurate to malformed teeth,and it has great application potential in computer-assisted orthodontic treatment diagnosis.