Data fusion and smoothing for probabilistic tracking of viral structures in fluorescence microscopy images

Abstract

Automatic tracking of viral structures displayed as small spots in fluorescence microscopy images is an important task to determine quantitative information about cellular processes. We introduce a novel probabilistic approach for tracking multiple particles based on multi-sensor data fusion and Bayesian smoothing methods. The approach exploits multiple measurements as in a particle filter, both detection-based measurements and prediction-based measurements from a Kalman filter using probabilistic data association with elliptical sampling. Compared to previous probabilistic tracking methods, our approach exploits separate uncertainties for the detection-based and prediction-based measurements, and integrates them by a sequential multi-sensor data fusion method. In addition, information from both past and future time points is taken into account by a Bayesian smoothing method in conjunction with the covariance intersection algorithm for data fusion. Also, motion information based on displacements is used to improve correspondence finding. Our approach has been evaluated on data of the Particle Tracking Challenge and yielded state-of-the-art results or outperformed previous approaches. We also applied our approach to challenging time-lapse fluorescence microscopy data of human immunodeficiency virus type 1 and hepatitis C virus proteins acquired with different types of microscopes and spatial-temporal resolutions. It turned out, that our approach outperforms existing methods.

Keywords: Bayesian sequential estimation; Biomedical imaging; Covariance intersection algorithm; Microscopy images; Multi-sensor data fusion; Particle tracking.