Fast Data Processing Methods For Super Resolution Localization Microscopy

Super-resolution localization microscopy (LM) achieves a high spatial resolution up to20nm, and has become an important tool for various biological applications by offering nanoscale resolution to reveal the complex working mechanisms of cellular machinery. But the temporal resolution of LM is greatly sacrificed to achieve such a high spatial resolution. Therefore, in the past years, great efforts have been carried out to improve the temporal resolution of LM by using faster cameras or employing high-density molecule localization methods. However, faster cameras will produce huge data flow and massive data, which will bring big challenge to the data transfer, data storage and data analysis. And, the widespread use of high-density localization methods are far mainly hindered by their slow data processing rates. To meet these challenges, here we develop several fast data processing methods basing on efficient data processing algorithms and high-performance computing platforms. The main works are summarized as follows.(1) Fast sparse molecule localization method. Utilizing the radial symmetry nature of the point spread function for a proper-aligned microscopy, we derive a fast and precise localization algorithm, termed maximum radial symmetry estimator (MrSE). The simulation and experimental results indicate that the localization precision of MrSE is close to the theoretical boundary, while the localization speed is over1000times faster than traditional sparse localization algorithms. By combining the MrSE-based data analysis method with the graphics processing unit (GPU) platform, the data processing speed is accelerated to650Mpix/s, which is capable of processing the data from sCMOS cameras in real-time.(2) High efficiency data reduction method. Utilizing the sparsity nature of the fluorescence molecules in LM, we derive a high efficiency data reduction algorithm. The simulation and experimental results indicate that raw data can be compressed～20times by this algorithm, while the final super-resolution image is not influenced. By combining this algorithm with the field programmable gate array (FPGA) platform in the camera, the data reduction speed is accelerated to532Mpix/s, which can significantly reduce the challenge of data transfer and data storage for LM imaging with sCMOS cameras.(3) Fast high-density molecule localization method. By performing a deoverlapping filter on the high-density molecule images, we derive a fast high-density localization algorithm, which can precisely localize the overlapping molecules. The simulation and experimental results indicate that this algorithm significantly improves the spatial resolution and temporal resolution of LM. Meanwhile, this algorithm exhibits comparable localization accuracy with DAOSTORM, while reducing the algorithm complexity with two orders of magnitudes. We further develop a dedicated processor based on this algorithm, termed POWERs, which can directly access the data form camera and perform high-density molecule localization independently. POWERs is capable of processing raw image data at200Mpix/s, which is far beyond the data acquisition rate of EMCCD cameras.In this dissertation, we present a series of fast data processing algorithms based on the inherent feature of raw images and optical system in LM. Then, we combine these algorithms with high-performance computing platforms to solve the challenges of data transfer, data storage and data analysis in LM. The presented methods can promote the wide spread use of fast super-resolution localization microscopy.