Fast 3D localization algorithm for high-density molecules based on multiple measurement vector compressed sensing
Fast 3D localization algorithm for high-density molecules based on multiple measurement vector compressed sensing
复制标题
基于多测量矢量压缩感知的高密度分子快速3D定位算法
DOI：
10.1016/j.optcom.2021.127563
复制
发表时间：
2021
影响因子：
2.4
通讯作者：
Li Wenguo
中科院分区：
文献类型：
--
作者：
Zhang Saiwen;Deng Yaqi;Lin Danying;Yu Bin;Chen Danni;Zhu Qiuxiang;Tian Yea;Wu Jingjing;Guangfu Zhang;Wen Bing;Li Wenguo
单分子定位显微镜（SMLM）基于一系列稀疏分布分子的图像实现超分辨率，这使得其非常耗时。为了提高时间分辨率，一些高密度定位算法已被开发，计算速度被认为是评价其性能的重要因素之一。本研究首次构建了三维多测量向量压缩感知（3D-MMV-CS）模型，并采用多稀疏贝叶斯学习算法进行求解，实现了三维单分子定位，获得了三维超分辨荧光图像。3D-MMV-CS成功地从一系列模拟高密度荧光图像中恢复了3D超分辨图像。重建的计算时间仅随着原始图像数量的增加而略微增加，并且随着荧光分子密度的增加而大致保持恒定。仿真结果表明，当原始图像数达到1900时，3D-MMV-CS算法与L1-Homotopy算法和基于CVX的压缩感知STORM（CSSTORM）算法相比，前者的速度提高了100倍，后者的速度提高了5个数量级。这种出色的计算速度性能将允许3D-MMV-CS方法常规地适用于SMLM。利用基于几何学的三维SMLM系统的实验数据成功地对3D-MMV-CS方法进行了验证。
Single-molecule localization microscopy (SMLM) achieves super-resolution based on a series of images of sparsely distributed molecules, which makes it time-consuming. To improve the temporal resolution, several high-density localization algorithms have been developed, and the computational speed is considered one of the important factors for evaluating their performance. In this study, for the first time, a three-dimensional (3D) multiple measurement vector compressed sensing (3D-MMV-CS) model was constructed and then solved using a multiple sparse Bayesian learning algorithm to realize 3D single-molecule localization and obtain 3D super-resolved fluorescence images. The 3D-MMV-CS successfully recovered a 3D super-resolved image from a series of simulated high-density fluorescence images. The computation time for reconstruction only increases marginally with an increase in the number of raw images and approximately stays constant with an increase in the density of fluorescent molecules. Simulation results demonstrated that, when the number of raw images reaches 1900, compared with the other two algorithms L1-Homotopy and compressed sensing STORM (CSSTORM) with CVX, the 3D-MMV-CS method is 100 times faster than the former and five orders of magnitude faster than the latter. This outstanding performance in computation speed would allow the 3D-MMV-CS approach to be routinely applicable to SMLM. Validation test of the 3D-MMV-CS method was performed successfully with experimental data from an astigmatism-based 3D SMLM system.