【作者】 张晨；

【导师】 陈键伟； 黄振立；

【作者基本信息】 华中科技大学 ， 光学工程， 2023， 硕士

【摘要】 超分辨定位成像可以实现20～30nm的分辨率,为生物医学研究提供了新的技术支撑。超分辨定位成像是一种典型的计算光学成像技术,需要借助合适的单分子定位算法来处理大量的原始荧光图像,以获得足够数量的定位点来重建超分辨图像。在二维超分辨定位成像过程中,基于压缩感知的单分子定位算法可以处理高密度荧光分子图像,减少超分辨图像重建所需的原始图像帧数。但是,该方法的图像处理速度较慢,严重影响了超分辨图像重建的效率。另一方面,在三维超分辨定位成像中,基于像散的轴向位置编码方法被人广泛应用,但该方法的轴向定位精度较低。最后,大视野超分辨定位成像是当前的新需求,可以用于跨尺度细胞功能成像。但是,与该需求相配套的点云拼接算法较为匮乏。针对这三方面的问题,本文分别从荧光分子的横向定位、轴向定位以及点云拼接等三方面着手,着力于提升超分辨定位成像关键算法的性能,从而实现高质量的大视野超分辨图像重建。(1)提出了一种基于级联压缩感知的高密度分子定位算法(Lite-CSR)。该算法在级联压缩感知算法(CSR)的基础上,通过优化分级求解模型,实现了更快的定位速度和更高的定位精度。仿真测试表明,相比于CSR算法,Lite-CSR算法的定位精度平均提高了 1.6倍,定位速度平均提高了 20倍。实验测试表明,Lite-CSR算法重建图像的分辨率为CSR算法的1.5倍,而重建速度比CSR算法提高了 40倍。(2)提出了一种基于多标定曲线的轴向分子定位算法。通过仿真不同尺寸点光源的成像过程,建立了基于多根标定曲线的轴向定位算法,对不同尺寸点光源的成像光斑进行分别定位。仿真测试表明,相比于单根标定曲线,基于多根标定曲线的定位方法,在离焦±400 nm范围内,分子轴向定位精度平均提高了 1～4倍。实验测试表明,在离焦不同距离处,利用多根标定曲线重建的三维超分辨图像,其轴向分辨率比利用单根标定曲线重建的图像提高了 1～2倍。(3)建立了一种适用于超分辨定位成像点云数据的拼接优化方法。该方法利用点云配准结果来建立不同视场之间的约束关系,通过非线性优化算法提升了拼接精度。仿真测试表明,相比于直接利用位移台位置信息的拼接算法,新方法的二维超分辨拼接精度提高了 4倍,三维超分辨拼接精度提高了 10倍。实验测试表明,在二维超分辨图像中,利用该方法获得的拼接图像比直接利用位移台信息拼接的图像更接近于原始图像,前者跟原始图像的结构相似度为0.78,而后者跟原始图像的结构相似度为0.66。在三维图像情形下也观察到了类似的结果。更多还原

【Abstract】 Super-resolution localization microscopy(SRLM)is able to provide a resolution as high as 20～30 nm,and thus offers new technical support for biomedical researches.SRLM is a typical computational optical microscopy technology,which requires the use of a suitable single-molecule localization algorithm to process a large number of raw fluorescence images and obtain a sufficient number of localizations to reconstruct a superresolution image.In two-dimensional SRLM,the single-molecule localization algorithm based on compressed sensing can process high-density fluorescence images,and thus reduces the number of raw image frames required for reconstructing a super-resolution image.However,the image processing speed of this algorithm is slow,which seriously affects the efficiency of super-resolution image reconstruction.In 3D SRLM,the axial position encoding method based on astigmatism is widely used.However,affected by various factors,the axial localization accuracy of this method is low.On the other hand,SRLM with large field of view(FOV)is a new demand,which can be used to realize crossscale imaging of cellular functions.However,the required point cloud stitching algorithm is relatively scarce.This thesis tried to improve the performance of key algorithms in SRLM,and focused on three aspects:lateral localization,axial localization,and point cloud stitching of fluorescent molecules,with the goal of achieving high-quality and large FOV in super-resolution image reconstruction.(1)A high-density molecule localization algorithm based on cascaded compressed sensing was proposed(Lite-CSR).Based on the cascaded compressed sensing algorithm(CSR),this algorithm optimizes the hierarchical solution model and achieves a faster localization speed and a higher localization accuracy.Simulation tests showed that,compared with the CSR algorithm,the localization accuracy of the Lite-CSR algorithm is average increased by 1.6 times,and the localization speed is average increased by 20 times.Experimental tests showed that the resolution of the images reconstructed by the Lite-CSR algorithm is 1.5 times that of the CSR algorithm,and the reconstruction time is only one fortieth of the CSR algorithm.(2)A axial molecule localization algorithm based on multiple calibration curves was proposed.The imaging process of point sources with different sizes were simulated,and a new molecule localization algorithm based on multiple calibration curves was established to improve axial localization accuracy.Simulation tests showed that,the axial localization accuracy of the new localization algorithm based on multiple calibration curves is 1～4 times higher than that from a single calibration curve in the defocus range of ± 400 nm.Experimental tests demonstrated that the axial resolution of the image reconstructed from multiple calibration curves at different defocus distances is 1～2 times higher than that reconstructed from a single calibration curve.(3)A point cloud stitching optimization method suitable for super-resolution data has been established.This method uses point cloud registration results to establish the constraint relationship among different FOV,and improves the stitching accuracy through a nonlinear optimization algorithm.Simulation tests showed that,compared with the stitching algorithm that directly uses the positions of the translational stage,the accuracy of the new stitching method is increased by 4 times in two-dimensional super-resolution images,and 10 times in three-dimensional super-resolution images.Experimental tests showed that,in twodimensional super-resolution images,the stitched image obtained by this new method is notably closer to the original image than the image stitched from translational stage positions,because the former has a structural similarity of 0.78 with the original image,while the latter has a structural similarity of 0.66 with the original image.Similar results were found in three-dimensional super-resolution images.更多还原