Research On Crack Detection Method Of Cracked Teeth Based On Deep Learnin

Cracked tooth is a dental disease which is difficult to diagnose,and is often accompanied by some clinical symptoms such as chewing pain,swollen gums,and cold irritation.Excessive occlusal forces,materials used for dental restorations,and structural weaknesses of the teeth are the factors which contribute to the occurrence of the cracked tooth.Today,cracked tooth have become the third most common cause of tooth loss in adults,after periodontal disease and dental caries.The early diagnosis of cracked tooth and timely restorative intervention can effectively limit the development of cracks.Currently,it is difficult to diagnose the cracked tooth,The methods to diagnose the cracked tooth are based on conventional detection and medical image-based crack detection.However,due to the influence of visual fatigue during the long diagnosis process,clinicians may be unable to accurately determine the presence of cracks on the teeth,which may continue to aggravate the condition if patients with cracked tooth do not receive timely treatment.In view of that deep learning has made great breakthroughs in the detection of other diseases in the oral field and the detection of cracks in the engineering field.In order to provide clinicians with a more intelligent solution for diagnosing cracked tooth,this paper adopts deep learning-based methods to complete the detection of cracks on teeth,and the main research contents are as follows:(1)A method based on Res Net combined with sliding windows is proposed to achieve the detection of cracks on the teeth.The pretrained Res Net is first fine-tuned to suit the tooth crack binary classification task.The Res Net networks with three different layers were trained and validated on a self-prepared tooth crack binary classification dataset,and the Res Net50 was finally selected as the tooth crack classifier based on the results on the test set.The classifier was then tested on 100 tooth crack images with a resolution of 1920 × 1080 pixels through a sliding window method,and the final average accuracy obtained was 90.39%.The experimental results show that the method can effectively detect cracks on teeth in various situations,which can provide clinicians with a smarter diagnostic solution in diagnosing cracked tooth in the clinic.(2)A method based on Deep Labv3+ is proposed to achieve the detection and segmentation of cracks on the teeth.By training and testing the models on the tooth crack segmentation dataset established in this section using the transfer learning method,the effect of Deep Labv3+with different backbone networks on the accuracy of tooth crack segmentation is investigated,and further compared with FCN,Seg Net and U-Net.Based on the testing results of each model,the best performing Res Net50-Deep Labv3+ model is selected as the base model architecture for the subsequent study.Besides,the reasons for the poor segmentation of some tooth crack images are summarized and analyzed to provide the basis for the optimization of the semantic segmentation models.(3)In order to further improve the segmentation accuracy of tooth cracks under various complex situations,the Deep Labv3+ model with the backbone network of Res Net50 is improved in this section.Firstly,the feature pyramid network was used to fuse the feature maps of different layers in the first three stages of the Res Net50 to be used as the input of the decoding layer,so as to increase the feature information.Then the atrous convolutional layers connected in parallel in the ASPP module of the original Deep Labv3+ model is changed to a densely connected way,which enables the model to obtain denser pixel sampling and larger receptive field,thus enhancing the feature extraction capability of the model.Finally,a BAM attention mechanism branch is added to the feature map from the fourth stage in the Res Net50,and then it is fused with the feature map processed by the densely connected ASPP module to extract rich contextual information.The experimental results showed that the improved model optimizes the segmentation of tooth cracks under various complex conditions.The improved Deep Labv3+ model achieved MPA and MIo U with a value of 75.52% and 75.07% on the test dataset,which are 2.85% and 3.01% better than the original Deep Labv3+ model.