【作者】 刘宇波；

【导师】 杨萃；

【作者基本信息】 华南理工大学 ， 通信与信息系统， 2019， 硕士

【摘要】 脊柱侧凸是严重危害青少年身心健康的脊柱畸形生长疾病,早期矫正、持续监测是治疗的关键。目前主要使用X射线对该疾病进行检测,但X射线具有辐射性,不利于病情的持续监测。三维超声能通过扫描脊柱获得脊柱的三维结构图,具有无辐射、易操作、低成本等特点,是一种替代X射线的方案。在三维超声的多种扫描方式中,使用机械装置控制超声探头运动的机械自动扫描具有精确定位、全自动、可重复的特点,适合于脊柱侧凸的检测。在基于机械自动扫描的三维超声自动扫描系统中,精确的脊柱扫描路径规划是获得准确的脊柱三维结构图的关键。通常可以利用摄像机等设备获取探头前方的彩色图像信息,然后利用图像分割算法对彩色图像进行分割提取脊柱目标区域,最后在脊柱目标区域内规划出最优扫描路径,完成脊柱扫描路径的规划。随着深度学习的不断发展,基于卷积神经网络的图像分割算法在特征提取、分割准确率上相比于传统方法有更好的表现,因此,本文基于卷积神经网络对脊柱三维超声自动扫描系统做了以下研究:(1)在脊柱扫描路径规划算法中,搭建了FCN、U-Net和Deeplab三种卷积神经网络,对比了三种网络对深度摄像仪拍摄的彩色图分割脊柱区域的效果,同时结合深度图的多视角信息,使用四通道法和双特征提取网络结构方法对深度摄像仪拍摄的深度图、彩色图进行整合,提高分割的准确率。最后使用基于人体轮廓的分割结果判定方法和最小二乘曲线拟合方法对分割结果进行后处理,得到较为精确的脊柱扫描曲线。(2)借鉴移动最小二乘形变方法提出以脊柱线为控制曲线对人体轮廓进行变形的数据增强方法,拟合多种脊柱侧凸情况,增加数据集的多样性,从而提高网络的泛化能力。(3)以医学超声仪器Sonix RP、六自由度机器人、深度摄像仪Kinect、工控计算机为基础搭建了一套脊柱三维超声自动扫描系统,在机械自动控制程序中嵌入改进的脊柱扫描路径规划算法,实现了脊柱的精确扫描与成像。更多还原

【Abstract】 Scoliosis is a spinal deformity disease that seriously harms the physical and mental health of adolescents.Early correction and continuous monitoring are the key to treatment.The mainly way to detect Scoliosis is X-ray,but X-rays are radiation-sensitive,which is not conducive to continuous monitoring of the disease.Three-dimensional ultrasound can obtain the three-dimensional structure of the spine by scanning the spine.It has the characteristics of no radiation,easy operation,low cost,etc,which is an alternative to X-ray.In the various scanning modes of three-dimensional ultrasound,the mechanical automatic scanning using the mechanical device to control the movement of the ultrasonic probe has the characteristics of precise positioning,full automatic and repeatability,and is suitable for the detection of scoliosis.In a three-dimensional ultrasound automatic scanning system based on mechanical automatic scanning,accurate spinal scan path planning is the key to obtaining accurate three-dimensional structure of the spine.Generally,to complete the planning of the spine scan path,we can obtain the color image information in front of the probe by using a device such as a camera,and then use the image segmentation algorithm to segment the color image to extract the target region of the spine,and finally plan the optimal scan path in the target region of the spine.With the continuous development of deep learning,the image segmentation algorithm based on convolutional neural network has better performance in feature extraction and segmentation accuracy than traditional methods.Therefore,based on convolutional neural network,the paper did the following research for the three-dimensional ultrasound automatic scanning system:(1)In the spinal scan path planning algorithm,three convolutional neural networks which are FCN,U-Net and Deeplab,were built to compare the effects of three networks on segmenting the spine region in the back view of color map taken by depth camera.At the same time,combined with the multi-view information of the depth map,the four-channel method and the dual feature extraction network structure method was used to integrate the color map and the depth map information taken by depth camera to improve the segmentation accuracy.Finally,the segmentation result is post-processed using the segmentation result determination method based on human contour and the least squares curve fitting method to obtain a more accurate spine scan curve.(2)Using the moving least squares deformation method to propose a data enhancement method for deforming the contour of the human body with the spinal line as the control curve,fitting a variety of scoliosis and increasing the diversity of the data set,thereby improving the generalization ability of the network..(3)Based on the medical ultrasound instrument Sonix RP,six-degree-of-freedom robot,depth camera Kinect,industrial computer,a three-dimensional ultrasound automatic scanning system was built.With an improved spinal scanning path planning algorithm embedded into the mechanical automatic control program,accurate scanning and imaging of the spine is achieved.更多还原