Biases in particle localization algorithms.
Automated particle locating algorithms have revolutionized microscopy image analysis, enabling researchers to rapidly locate many particles to within a few pixels in a microscope image. The vast majority of these algorithms operate through heuristic approaches inspired by computer vision, such as identifying particles with a blob detection. While rapid, these algorithms are plagued by biases [4, 15, 24], and many researchers still frequently ignore or understate these biases. In this paper, we examine sources of biases in particle localization. Rather than exhaustively examine all possible sources of bias, we illustrate their scale, the large number of sources, and the difficulty of correcting the biases with a heuristic method. We do this by generating a series of simple images, introducing sources of bias one at a time. Using these images, we examine the performance of two heuristic algorithms throughout the process: a centroid algorithm and a Gaussian fitting algorithm. We contrast the two heuristic methods with a new approach based on reconstructing an image with a generative model to fit the data (Parameter Extraction from Reconstructing Images, or PERI). While the heuristic approaches produce considerable biases even on unrealistically simple images, the reconstruction-based approach accurately measures particle positions even in complex, highly realistic images. We close by reiterating the fundamental reason that a reconstruction-based approach accurately extracts particle positions -- any imperfections in the fit both demonstrate which sources of systematic error are still present and provide a roadmap to incorporating them.
Publisher URL: http://arxiv.org/abs/1801.03581