| The purpose of dental area semantic segmentation is to automatically segment the upper and lower teeth after inputting patient photos,deep learning-based tooth area segmentation methods can be affected by the low ratio of foreground area,insufficient feature recognition of different tooth sizes and tooth edges at present,which limits its applicability.In this paper,I propose a two-stage deep learning-based dental image segmentation system,which includes a Transformer-based face detector,and a Deeplab V3+ based algorithm,which can automatically detect patients’ face and perform semantic segmentation of teeth.The Transformer-based face detection network in this system is based on DAB-DETR.Previous methods have used location information but ignored the scale of object,by combining area and content information,I propose a method can explicitly and dynamically modeling anchor boxes in the query generation module,effectively improve the accuracy of the algorithm.Algorithm in this paper evaluated on self-built database,and the AP(Average Precision)reached 0.9168.In addition,I designed a Deeplab V3+ based encoder-decoder network to achieve automatic semantic segmentation of tooth areas.Specifically,in this paper,the refined algorithm added a feature fusion module to integrate different levels of features,and improve the segmentation of different sizes of tooth areas,and a composite loss combining the mutual information of pixels in the picture is designed to improve the segmentation of tooth edge.Algorithm in this paper evaluated on self-built database,the MIo U(Mean Intersection over Union)reached 0.9104,the MPA(Mean Pixel Accuracy)reached 0.9543,and the average inference time per picture is less than 13 ms. |
