Root Mean Square Minimum Distance as a Quality Metric for Stochastic Optical Localization Nanoscopy Images

A localization algorithm in stochastic optical localization nanoscopy plays an important role in obtaining a high-quality image. A universal and objective metric is crucial and necessary to evaluate qualities of nanoscopy images and performances of localization algorithms. In this paper, we propose root mean square minimum distance (RMSMD) as a quality metric for localization nanoscopy images. RMSMD measures an average, local, and mutual fitness between two sets of points. Its properties common to a distance metric as well as unique to itself are presented. The ambiguity, discontinuity, and inappropriateness of the metrics of accuracy, precision, recall, and Jaccard index, which are currently used in the literature, are analyzed. A numerical example demonstrates the advantages of RMSMD over the four existing metrics that fail to distinguish qualities of different nanoscopy images in certain conditions. The unbiased Gaussian estimator that achieves the Fisher information and Cramer-Rao lower bound (CRLB) of a single data frame is proposed to benchmark the quality of localization nanoscopy images and the performance of localization algorithms. The information-achieving estimator is simulated in an example and the result demonstrates the superior sensitivity of RMSMD over the other four metrics. As a universal and objective metric, RMSMD can be broadly employed in various applications to measure the mutual fitness of two sets of points.