Issues in the Benchmarking of Image Analysis Algorithms for Superresolution Microscopy
Authors:Shane Stahlheber, Shane Stahlheber
Mentor:Alexander Small, Professor of Physics, California State Polytechnic University Pomona
Superresolution localization microscopy requires accurate and precise localization algorithms. We have developed a plugin for ImageJ, called M2LE, which can localize molecules quickly and distinguish between single-molecule and multiple-molecule images using a shape test that requires only a single iteration. Localization is accomplished via a fast maximum-likelihood algorithm that uses the separable property of the Gaussian to independently fit two 1-D Gaussians along the x- and y-directions. To assess the performance of M2LE, we tested the plugin with realistic simulated images of single and multiple molecule images. We first found the optimal shape test parameters that accept most single-molecule images, and then the optimal signal-to-noise cutoff parameter for identifying potential molecules from noise. These two parameters have the greatest impact on what parts of the image go on to be analyzed. Using these optimal parameters, we then assessed (1) the tendency of the algorithm to find molecules from the tail of a point-spread function in high signal-to-noise cases, (2) the effects of regions-of-interest size and overlap tolerances, (3) the ability of shape tests to identify multi-molecule images as a function of molecular separation and ratio of photon counts from two molecules, and (4) the performance of the entire process—the number of molecules identified and their corresponding localization precision and accuracy. The parameters for the M2LE plugin were optimized to maximize single-molecule image acceptance rates while minimizing multi-molecule and zero-molecule acceptance rates. These methods and results can be used to identify the optimal M2LE parameters to use for experiments, as well as to compare the performance with other localization microscopy software.