A Quantitative Analysis on Two RFS-Based Filtering Methods for Multicell Tracking

Multiobject filters developed from the theory of random finite sets (RFS) have recently become well-known methods for solving multiobject tracking problem. In this paper, we present two RFS-based filtering methods, Gaussian mixture probability hypothesis density (GM-PHD) filter and multi-Bernoulli filter, to quantitatively analyze their performance on tracking multiple cells in a series of low-contrast image sequences. The GM-PHD filter, under linear Gaussian assumptions on the cell dynamics and birth process, applies the PHD recursion to propagate the posterior intensity in an analytic form, while the multi-Bernoulli filter estimates the multitarget posterior density through propagating the parameters of a multi-Bernoulli RFS that approximates the posterior density of multitarget RFS. Numerous performance comparisons between the two RFS-based methods are carried out on two real cell images sequences and demonstrate that both yield satisfactory results that are in good agreement with manual tracking method. 1. Introduction In previous related research in the field of medicine and biology, people often use manual visual inspection to observe living cells behaviors, such as the moving velocity and the density of cell population. Manual analysis methods always are limited by many challenges including large volume of image data and tedious manual operations. Therefore, it is of great significance to find an automatic and reliable way to track multiple cells. Over the last decades, many automatic tracking methods have been developed with rapid development of processing and computer vision technologies. Furthermore, numerous applications have been found for analyzing time-lapse cell microscopy imagery including fluorescent image [1, 2] and nonstaining cell image [3]. In the field of cell tracking, automatic and accurate tracking needs to overcome problems that mainly stem from two factors: cell factors such as population size, mitosis, and shape variability and imaging factors such as level of noise, image resolution ratio, and image contrast. The main objective of cell motion analysis is to determine the number of cells and the state of each cell, and visual multicell tracking methods can be divided into two categories, deterministic methods and stochastic methods. Deterministic methods usually handle the detections and tracking tasks separately [4–7], but tracking may fail under problematic imaging conditions such as large cell density, cell division events, or segmentation errors. Stochastic methods mostly fall in the category of Bayesian framework, and