Detection And Classification Of Dental Caries Based On Uncertain Aware And Structured Segmentation

The caries detection and classification method based on panoramic images uses computer vision technology to automatically detect and classify dental caries lesions,so as to segment caries areas and classify the relative category according to the degree of lesion erosion.For detecting lesion areas from a 2D image with classification,object segmentation network is generally used to segment the lesion area along with using end-to-end methods to classify the detected lesions directly.However,there are two problems: 1)With surface projection reconstruction of all teeth in the panoramic images,a large amount of noise and artifacts are introduced,Artifacts performing extremely similar characteristics with lesions therefore cause serious interference on the actual detection of caries,At the same time,the diversity of scale and morphology of caries also poses a significant challenge to forming stable caries features.2)Due to neural networks’ insufficient ability to exactly locate caries boundary,the classifier cannot accurately capture the depth of caries erosion.Without erosion depth features ultimately leads to unreliable classification results during the diagnostic process.For the above two problems,this article has carried out the following research:(1)For problem 1,this thesis proposes a novel semi-supervised framework based on multi-level uncertainty awareness.Three different types of interference are generated by adding noise to panoramic slices,making iterations differences,and implementing multi-scale predictions.Meanwhile,deep-supervision training is applied on a multi-layer decoder,constraining the decoding layers to produce as consistent predictions as possible,in order to generate a number of caries prediction samples.Finally,Monte Carlo is used to sampling the prediction results formed under various interferences uniformly to integrate a more robust uncertainty map.Based on this map,the inter-class feature differences between the actual lesions and artifacts are determined.(2)For problem 2,this thesis proposes a method for tooth instance reconstruction as well as segmentation based on structured awareness.The method performs a more granular structural segmentation of the tooth’s internal structure.By introducing image-level features and integrating them into high-level semantic features of the tooth instance,it supplements more granular tooth structure information for more reasonable structural segmentation.Then,a structural awareness module is designed to distinguish between structural masks and its corresponding predictions,entail the structural segmentation network produce a rather complete tooth structure.In addition,this thesis proposes an image repair method based on the diffusion model to reconstruct abnormal regions in tooth instances that destroy the tooth structure,enabling them to generate typical tooth structures in structured segmentation.(3)Using method(1)to detect caries lesions from panoramic images and then using method(2)to segment the tooth structure,the erosion degree line employed to classify caries lesions in the internal structure of the tooth can be obtained through the geometric solution of the image,and the location of detected caries within different grading polygons can be determined to make more accurate caries classification ultimately.