Study On Mandible Segmentation In The Temporomandibular Joint Cone Beam CT Image

Temporomandibular joint(TMJ)is one of the most complex joints in human body.Temporomandibular joint disorder(TMD)is a common disease with high prevalence and closely related to multi-disciplines,which has attracted more and more attention from stomatologists.Cone beam computed tomography(CBCT)is regarded as the main medical imaging tool to diagnose the changes of TMJ condylar joints.The segmentation of TMJ bone structure in CBCT image can be applied to the 3D quantitative measurement of condylar bone remodeling,which has important clinical application value for the diagnosis and treatment of TMD diseases.Due to the large individual differences,low image contrast and large amount of 3D data,there are still great challenges in the accurate segmentation of TMJ mandible in CBCT images.This thesis focuses on the study of TMJ CBCT image segmentation algorithm based on deep learning under the limited GPU memory.(1)To demonstrate the feasibility of deep learning method in TMJ CBCT image segmentation,a mandibular segmentation algorithm of CBCT image based on global down-sampling(Seg Net-GDS)is proposed.Global downsampling is used to reduce the amount of network input data,so that the U-Net segmentation network can complete network training under limited GPU memory.Compared with the classical segmentation methods,clinical images results show that our method achieves more accurate results,and the average dice similarity coefficient is 97.60%.(2)Aiming at the loss of image detail information caused by global down-sampling,a multi-stage image pyramid network(MIP-Net)for the CBCT image mandibular segmentation is proposed.The model adopts the characteristic that the resolution of image pyramid increases gradually from top to bottom,we design several feature fusion approaches to fully integrate the low-resolution coarse segmentation results with the high-resolution feature maps,so as to realize the accurate segmentation of mandible from coarse to fine in CBCT images.Clinical images results show that our method yeild better segmentation performance than the global down-sampling method based on U-Net,and the average dice similarity coefficient increases to 98.04%.(3)To further improve the segmentation performance of the MIP-Net,a multi-view ensemble learning network(MEL-Net)for CBCT image mandibular segmentation is proposed.Anisotropic multi-view resampling provides multiple weak learners with different spatial semantic information to output reasoning results with different preferences through multiple weak learners,then we utilize the strong learning network to synthesize more comprehensive context information to realize accurate segmentation.The experimental results show that the MEL-Net has stronger segmentation performance than the MIP-Net network,and the average dice similarity coefficient increases to 98.18%.This thesis is devoted to the study of high-resolution CBCT image segmentation algorithm under limited GPU memory,and uses deep learning to solve difficult problems in clinical medical image applications.Our research results have good clinical application prospects,and provide new ideas for the research of 3D medical image segmentation as well as other image processing algorithms.