脊髓是人体中枢神经系统的重要组成部分。研究表明，脊髓受到肿瘤侵袭时会严重损害人体感知、运动等功能。治疗脊髓肿瘤最为有效的方式是手术切除，准确识别脊髓肿瘤边界对手术切除引导具有重要意义。超声定位显微成像（ultrasound localization microscopy，ULM）基于超声微泡造影剂可以突破传统超声成像衍射极限，对微血管进行准确有效的成像。本研究旨在进行ULM成像方法研究，并将其应用于脊髓肿瘤，主要包含以下三项工作：（1）运动矫正研究。数据采集过程中脊髓受到呼吸、心跳等运动的影响，本研究采用基于归一化互相关（normalized cross-correlation，NCC）的非刚性运动矫正算法有效消除了超声定位显微图像的运动伪影。与刚性运动矫正算法相比，尽管非刚性运动矫正算法时间复杂度更高，但运动矫正效果更好，矫正后的图像中血管半高全宽（full width at half maximum，FWHM）更小，分辨率更高。进一步地，针对基于NCC的非刚性运动矫正算法时间复杂度过高的问题，本研究将基于深度学习的运动估计算法用于ULM运动矫正。实验结果表明，基于深度学习的运动估计算法可以准确进行非刚性运动估计，并极大缩短运动估计时间。（2）微泡定位方法研究。针对ULM成像中采用单一深度微泡拟合得到的点扩散函数（point spread function，PSF）进行微泡定位不准确的问题，本研究提出了使用不同深度微泡拟合得到的PSF加权平均的结果进行微泡定位的方法。仿真结果表明，本研究提出方法的微泡定位灵敏度更高。（3）基于ULM的脊髓肿瘤应用探索研究。为验证ULM在人体脊髓应用的可行性，本研究通过采集术中人体脊髓海绵状血管瘤超声B模式以及谐波模式数据，结合上述非刚性运动矫正算法和基于深度加权平均PSF的微泡定位方法，实现了人体脊髓肿瘤ULM成像，并进一步提取ULM图像中血管密度、血流最大速度、最小速度、方向熵等定量指标。结果显示，正常脊髓区域血管密度高于海绵状血管瘤区域。本研究进行了基于NCC的非刚性运动矫正算法研究，并通过深度学习进行算法加速，提出了深度加权平均PSF的微泡定位算法，得到了更优的ULM成像效果。并通过在体数据验证了ULM对脊髓肿瘤边界识别的应用价值，为ULM成像的临床应用奠定了技术基础。
The spinal cord (SC) is an important part of the central nervous system. Studies have shown that spinal tumor (ST) invasion can seriously damage human sensory, motor and other functions. The most effective way to treat spinal tumor (ST) is surgical resection. Accurate identification of ST boundary is of great significance for surgical resection guidance. Ultrasound localization microscopy (ULM) based on ultrasound microbubble contrast agent can break through the diffraction limit of traditional ultrasound imaging and accurately and effectively image microvessels. The purpose of this study is to research ULM and apply it to ST, mainly including the following three work:(1) Motion correction (MoCo) research. During data collection, SC is affected by respiration, heartbeat and other movements. In this study, normalized cross-correlation (NCC) based non-rigid MoCo algorithm is used to effectively eliminate the motion artifacts in ultrasound localization microscopy images. Compared with the rigid MoCo, although the non-rigid MoCo has higher time complexity, the performance is better, and the corrected image has lower full width at half maximum (FWHM), i.e. higher resolution. Furthermore, aiming at the problem of high time complexity of non-rigid MoCo based on NCC, this study applies motion estimation algorithm based on deep learning to ULM MoCo. Experimental results show that the motion estimation algorithm based on deep learning can accurately perform non-rigid motion estimation and greatly shorten the time of motion estimation.(2) Study on microbubble localization method. In order to solve the problem that the point spread function (PSF) obtained by single depth microbubble fitting may not perform well in ULM, this study proposes a method of microbubble localization using weighted average PSF obtained by different depth microbubble fitting. Simulation results show that the proposed method is more sensitive to microbubble localization.(3) Research on the application of ST based on ULM. In order to verify the feasibility of ULM application in SC, in this study, ULM of human ST is achieved using intraoperative ultrasound B-mode and harmonic imaging data of cavernous hemangioma, combined with the above non-rigid MoCo and microbubble localization method based on depth weighted average PSF. Furthermore, quantitative indexes such as vascular density, maximum velocity, minimum velocity and direction entropy are extracted from ULM images. The results show that vascular density was higher in the normal SC than in the cavernous hemangioma.In this study, the non-rigid MoCo based on NCC is studied, and the algorithm is accelerated by deep learning. The microbubble localization algorithm of depth weighted average PSF is proposed, and better ULM imaging effect is obtained. The application value of ULM in boundary recognition of ST is verified by in vivo data, which lays a technical foundation for the clinical application of ULM.
