Deep Learning Methods for Detection and Tracking of Particles in Fluorescence Microscopy Images

Studying the dynamics of sub-cellular structures such as receptors, filaments, and vesicles is a prerequisite for investigating cellular processes at the molecular level. In addition, it is important to characterize the dynamic behavior of virus structures to gain a better understanding of infection mechanisms and to develop novel drugs. To investigate the dynamics of fluorescently labeled sub-cellular and viral structures, time-lapse fluorescence microscopy is the most often used imaging technique. Due to the limited spatial resolution of microscopes caused by diffraction, these very small structures appear as bright, blurred spots, denoted as particles, in microscopy images. To draw statistically meaningful biological conclusions, a large number of such particles need to be analyzed. However, since manual analysis of fluorescent particles is very time consuming, fully automated computer-based methods are indispensable. We introduce novel deep learning methods for detection and tracking of multiple particles in fluorescence microscopy images. We propose a particle detection method based on a convolutional neural network which performs image-to-image mapping by density map regression and uses the adaptive wing loss. For particle tracking, we present a recurrent neural network that exploits past and future information in both forward and backward direction. Assignment probabilities across multiple detections as well as the probabilities for missing detections are computed jointly. To resolve tracking ambiguities using future information, several track hypotheses are propagated to later time points. In addition, we developed a novel probabilistic deep learning method for particle tracking, which is based on a recurrent neural network mimicking classical Bayesian filtering. The method includes both aleatoric and epistemic uncertainty, and provides valuable information about the reliability of the computed trajectories. Short and long-term temporal dependencies of individual object dynamics are exploited for state prediction, and assigned detections are used to update the predicted states. Moreover, we developed a convolutional Long Short-Term Memory neural network for combined particle tracking and colocalization analysis in two-channel microscopy image sequences. The network determines colocalization probabilities, and colocalization information is exploited to improve tracking. Short and long-term temporal dependencies of object motion as well as image intensities are taken into account to compute assignment probabilities jointly across multiple detections. We also introduce a deep learning method for probabilistic particle detection and tracking. For particle detection, temporal information is integrated to regress a density map and determine sub-pixel particle positions. For tracking, a fully Bayesian neural network is presented that mimics classical Bayesian filtering and takes into account both aleatoric and epistemic uncertainty. Uncertainty information of individual particle detections is considered. Network training for the developed deep learning-based particle tracking methods relies only on synthetic data, avoiding the need of time-consuming manual annotation. We performed an extensive evaluation of our methods based on image data of the Particle Tracking Challenge as well as on fluorescence microscopy images displaying virus proteins of HCV and HIV, chromatin structures, and cell-surface receptors. It turned out that the methods outperform previous methods