Research On Tooth Image Segmentation Methods Based On Deep Learning

Tooth image segmentation is the key technology of medical image processing and analysis,and it is also the first step of computer-aided diagnosis.It is widely applied in the fields of orthodontic treatment,dental disease diagnosis,dental training system and corpse identification.This thesis focuses on the problem of fuzzy boundary edge in dental panoramic X-ray image and the problem of tooth image segmentation that is difficult to balance accuracy and real-time.The detailed research work is as follows:Aiming at the problem of fuzzy boundary edges,this thesis proposes a dental panoramic X-ray image segmentation method based on multi-scale location perception.To realize the feature sharing between features of different scales,this thesis redesigns the feature transfer mode of the spatial pyramid pooling group within the architecture.A location perception module that is extremely sensitive to pixel positions is embedded in each pooling branch to accurately locate the tooth region.This module expands the receptive field inside the network from a local area to an entire image,so that detailed information and global semantic information can interact with each other.In addition,focusing on the perception characteristics of the human visual system for edge information,the architecture uses multi-scale structural similarity to construct a boundary loss function.In this way,the network can strengthen its ability of reducing boundary loss via assigning higher loss to the boundary.Experimental results show that the multi-scale location perception method outperforms existing methods,and improves the tooth edge accuracy of segmentation.Tooth segmentation method is improved by above method for the first time based on multi-scale structural similarity,and achieves prominent performance.However,its architecture is relatively complicated in scale and also cannot effectively make full use of the contrast information that affects the sharpness of the edges.In response to this problem,a contrast assistance based dual-channel segmentation method for dental panoramic X-ray images is proposed.The architecture contains two channels: affinity channel and fuzzy enhancement channel.The affinity channel embedded with affinity module is used to calculate the affinity weights of the local semantics and the global semantics.In this way,the impact of global semantics on different local regions can be obtained to roughly segment the tooth area.The fuzzy enhancement channel obtains the potential mapping of the pixel positions and pixel intensities of the tooth edges to improve the characterization ability of the simple network,and at the same time refine the segmentation results of the affinity channel.Experiments compare and discuss the current mainstream method and the proposed dual-channel segmentation method.The results indicate the designed tooth segmentation network has comparatively high real-time and accuracy.