A Recurrent Neural Network for Particle Tracking in Microscopy Images Using Future Information, Track Hypotheses, and...

Area of Science:

• Cell biology
• Biophysics
• Image analysis

Background:

• Accurate particle tracking in time-lapse fluorescence microscopy is crucial for understanding subcellular and viral dynamics.
• Existing methods often struggle with complex trajectories, missing detections, and track initiation/termination.

Purpose of the Study:

• To develop a novel, automated particle tracking approach using deep recurrent neural networks.
• To improve the accuracy and robustness of particle tracking in challenging microscopy datasets.
• To eliminate the need for handcrafted features and manual training data.

Main Methods:

• A deep recurrent neural network architecture was employed, utilizing both forward and backward temporal information.
• The network jointly determines assignment probabilities, computes probabilities of missing detections, and estimates existence probabilities for track management.
• Track hypotheses are propagated to future time points to resolve ambiguities using later information.

Main Results:

• The proposed method demonstrated superior performance compared to previous approaches on benchmark datasets (Particle Tracking Challenge) and real microscopy data.
• The approach effectively handles track initiation and termination by determining existence probabilities.
• Eliminated the necessity for handcrafted similarity measures, motion features, and manually labeled training data.

Conclusions:

• The novel deep recurrent neural network-based particle tracking approach offers a significant advancement in analyzing dynamic biological processes.
• This method provides a robust and automated solution for particle tracking, applicable to both subcellular structures and virus dynamics.
• The approach's ability to learn without manual labels and handcrafted features makes it highly adaptable and efficient.