| Techniques such as image segmentation and landmarks detection of the spine can effectively improve the accuracy of spinal registration and reconstruction,thus reducing the error rate of spinal surgery.However,in clinical practice,physicians rely on manual methods to complete the segmentation of spinal images and detection of landmarks.Although traditional manual operation technology provides convenience for spinal diagnosis and therapy,there are many limitations such as operations are cumbersome and time-consuming,the need for experienced doctors to operate,and the results are affected by objective factors.To enhance the efficiency of treatment and upgrade the safety factor of spinal surgery,this thesis is dedicated to the segmentation of spinal magnetic resonance(MR)images and the automatic detection of three-dimensional landmarks of the spine.Based on deep learning and image processing technology,the following two studies have been completed respectively:In this thesis,an image segmentation module based on symmetric convolution channel is proposed to complete the segmentation of vertebral body blocks in spinal MR images.On the basis of constructing a symmetric convolution channel module,the network is equipped with other network modules.The proposed network gives full play to the advantages of cooperation between modules and improves the efficiency of network training.The main improvements include:1.The symmetric convolutional neural network proposed in the network provides the network with wider features extract the receptive field.This module also improves the network training performance in depth.2.To solve the loss of high-resolution feature maps caused by the deepening of the network,we add dynamic residual network.It can connect to the network layer to strengthen the transmission of network training parameters.3.We use the convolutional block attention mechanism to focus on feature graph extraction in image channel and space,which is more helpful to enhance the effectiveness of feature.It is effective in realizing model convergence.By comparing the existing FCN,U-Net,DeeplabV3+ and UNet++ networks,the Dice Similarity Coefficient(DSC)of spinal MR image segmentation have increased by 15.34%,7.08%,5.79%,and 3.1%,respectively.The results show that the improved network can improve the segmentation accuracy of spinal MR images in clinical practice.In spinal landmarks detection,we propose a three-dimensional landmarks detection method based on spinal geometric constraints.The two research contents are as follows:1.The three-dimensional convex hull detection algorithm is used to generate the threedimensional geometric convex hull nodes of the spine.The clustering analysis is performed on the three-dimensional convex hull nodes to obtain the contour landmarks of the spine.2.The proposed method automatically establishes the local system of the spine,and detect the geometric symmetry plane of the spine and the cross-section of key parts.Then,the method divides the three-dimensional model of the spine into three parts.Finally,the method extracts the spinal local landmarks in the segmented area.The average Euclidean distance difference between our method and the landmarks of expert group is 1.86±1.15 mm and 2.07±1.06 mm,respectively.The proposed method can complete the detection of 38 three-dimensional landmarks of the spine within the 30 s,which saves 5-6 times compared with the manual operation of experts.It will significantly reduce the time for spinal landmarks in the process of spinal diagnosis and treatment. |
