Three-dimensional Segmentation Of Human Temporal Bone Micro-organs Based On CT Images

The temporal bone is located in the inner ear of the human body.It is small in size and complex in structure.It contains not only bony structures such as the labyrinth,but also soft tissue structures such as nerves and blood vessels.Temporal bone CT examination is a necessary examination before inner ear surgery.It can help doctors find structural abnormalities of the inner ear and understand the pathological changes of the anatomical structure.However,only observing two-dimensional CT slices is not intuitive enough and is not conducive to image analysis.Automatically and accurately segment important anatomical structures of the temporal bone helps to accurately understand their mutual positional relationship and lesion damage,and assists doctors in planning surgery.This thesis conducts research on the three-dimensional segmentation technology of semicircular canal and facial nerve in temporal bone CT images.The area of the semicircular canal and facial nerve accounts for less than 1% of a single-layer CT slice,and they cannot be accurately segmented by traditional segmentation methods.Therefore this thesis has designed three-dimensional automatic segmentation algorithms based on deep learning technology,correspondingly.In order that the segmented model can better assist doctors in diagnosis and treatment,the segmentation model has been reconstructed based on 3D reconstruction technology,and an application has been developed to display the 3D model.The research has been mainly divided into three parts:(1)The first part focuses on the segmentation technology of semicircular canals.Aiming at the small proportion of semicircular canals in CT slices and the characteristics of fuzzy boundaries,the U-Net framework based on the encoder-decoder structure has been designed.A three-dimensional spatial attention mechanism module has been introduced in the encoder part of the original framework,and a three-dimensional channel attention mechanism module has been introduced in the decoder part of the network.The modules have improved the accuracy and robustness of algorithm segmentation.Experimental results have showed that the improved network achieves an average Dice coefficient of 92.5% in the test set.(2)The second part focuses on the 3D segmentation technology of facial nerve.In view of the characteristics of slender facial nerve structure,extremely narrow tube diameter,and the narrowest point occupying only seven or eight pixels in a single-slice CT,UNETR network with Transformer combined with Atrous spatial pyramid pooling module and Non-local self-attention module,have increased the efficiency of data transmission,reduced the amount of parameters,further expanded the ability to extract deep features and shallow features,and avoided the problem of missing segmentation that is prone to occur at the narrowest part of the facial nerve.The experimental results have showed that the average Dice coefficient is 81.53% on the test set.(3)In the third part,a temporal bone CT visualization system has been designed and implemented based on 3D reconstruction technology.This system has constructed the segmented temporal bone organs and provided a variety of interactive methods to help doctors diagnose and improved the application level of the algorithm.The three-dimensional reconstruction of temporal bone organs is of great significance for otologists to assist in pathological diagnosis,surgical planning,and guidance for repositioning manipulations of otolithiasis.