Image Classification And Segmentation Of Mandibular Impacted Wisdom Teeth Based On Deep Learnin

According to national and international literature,the extraction of obstructed mandibular third molars is difficult due to the high rate of obstruction(between 24% and 54%)and the limited operative space and unclear visualization.At this stage,clinicians need to use CBCT images for preoperative planning to reduce the risk and pain during the procedure.However,this approach is still limited to two-dimensional space,and there are still some discrepancies with the actual situation in three-dimensional space,which may lead to surgical accidents.With the help of the current hot deep learning technology,the computer can automatically identify multiple tissue structures(mainly 37,38,47,48 and mandibular canal tissues involved in extraction procedures)and types of obstructed teeth in CBCT images to visualize the spatial pattern of the relevant tissues.With the help of medical imaging software such as 3D Slicer,it is possible to measure the distance between certain tissues,the angle between them,and other parameters to assist the surgeon in diagnosis and treatment with greater precision.The main results of this study are as follows.(1)After exploring the relevant classification and segmentation dataset annotation methods in the medical community,we collected 1262 CBCT cases from Guiyang Stomatological Hospital and constructed the classification and segmentation datasets of CBCT images based on them,in which 207 CBCT annotations were completed for the classification task and 187 CBCT annotations were completed for the segmentation task.(2)In order to facilitate the later distribution and disclosure of the dataset,we converted the CBCT storage format from Dicom format to nii.gz format using the Monai framework,and saved the segmentation labels as 3D Slicer default format nrrd files and the classification labels as Excel tables.(3)Based on the first and second steps,we completed the training and inference process of deep learning-based image segmentation and classification models using frameworks such as Pytorch-Lightning and Monai,and saved the prediction results of the models as nii.gz files for easy visualization and display at a later stage.