Research On The Diagnosis Of Periodontitis In Panoramic Dental X-ray Film Based On Deep Learning

Periodontal diagnosis is the first step in dental treatment.The routine clinical examination method is to take panoramic dental X-rays.However,the diagnosis accuracy of general practitioners is low,and it is often difficult to detect and treat the lesions at an early stage.With the rapid development of deep learning methods in the field of medical image analysis,more and more studies have proved that this method has strong application potential in medical image-assisted diagnosis.In this thesis,the auxiliary diagnosis of periodontitis with panoramic dental film is studied based on deep learning method.The main research contents are summarized as follows:(1)To solve the problem of lack of diagnostic information in panorama and periodontitis,an image data set with diagnostic labeling is established.Specifically,images are firstly exported jointly with the Stomatological Hospital of Wuhan University to control the balance between healthy samples and periodontitis samples,the duplicate samples are eliminated by comparing the histograms,the size distribution of images is analyzed to retain the typical samples.Then,maxillofacial annotation rules and periodontitis annotation rules of panoramic images are developed jointly with periodontists,and deep learning Image annotation tool VIA(VGG Image Annotator)is selected for collaborative annotation.Finally,the maxillofacial areas of the screened 2602 panoramic films are labeled,and the degree of alveolar bone resorption and root bifurcation of each tooth in 1747 panoramic films are labeled.The established dataset is the largest maxillofacial segmentation and periodontitis diagnosis dataset to support the training and evaluation of deep learning models.(2)Aiming at the distortion of the external maxillofacial region interferes with the detection of tooth main object,according to the classical semantic segmentation Network structure,an EED-Net(Efficient Encoder-Decoder Network)model based on the Efficient semantic segmentation Network is proposed to achieve the binarization segmentation of the maxillofacial panorama.Specifically,the encoder based on residual structure is firstly used to learn the residual information of feature graph without introducing training parameters,so as to accelerate the model calculation speed.Secondly,a feature extractor with a variety of perceptual field sizes is designed to obtain different levels of depth semantic features.Finally,the number of decoder channels is the same as the categories of recognized objects,and the redundant information channels are discarded to simplify the output structure of the model.Experiments on panoramic slice maxillofacial segmentation show that,compared with the latest lightweight semantic segmentation model,the proposed model has fewer parameters,higher segmentation accuracy,and the output rate,which can meet the real-time application requirements.(3)To realize the judgment of the tooth position and periodontitis category in the panoramic film,a single-step target detection model based on FL-PDNet(Focal Loss Periodontitis Detection Network)is proposed.In terms of structural optimization,the proposed model reuses the convolution structure to improve the receptive field of the input connection layer.In the basic structure of the backbone network,the feature channels are combined first and then nonlinear activation is performed to reduce redundant parameters and improve the ability of shallow feature extraction.Channel pruning is performed on the target feature extraction structure,and the anchor point frame is established according to the tooth size distribution in the data set to improve the detection ability of the main target frame size.In terms of training enhancement,the proposed model combines EED-Net to perform zerovalue conversion on the non-maxillofacial region of the input image,and uses the focal loss function to amplify the detection error of difficult-to-classify samples and improve the model’s category detection accuracy.Experiments based on the panoramic film periodontitis detection data set show that the detection speed and accuracy performance of the proposed model are better than the original YOLOv4 model.In addition,based on the proposed model,the framework of the panoramic film periodontitis auxiliary diagnosis system is designed to realize the visualized automatic diagnosis from the dental film image to the periodontitis medical record.In summary,this article first establishes a panoramic film periodontitis detection data set,and then proposes an EED-Net model based on an efficient semantic segmentation network to extract the panoramic maxillofacial region,and thirdly proposes a single-step target detection model based on FL-PDNet to improve the jaw The detection accuracy of the diseased teeth in the facial area,and finally the application framework for the auxiliary diagnosis of periodontitis is established.