Deep probabilistic tracking of particles in fluorescence microscopy images

Area of Science:

• Biophysics
• Cell Biology
• Microscopy

Background:

• Accurate particle tracking in temporal fluorescence microscopy is crucial for quantifying dynamic biological processes.
• Existing methods often require manual parameter tuning and lack robust uncertainty quantification.

Purpose of the Study:

• To introduce a probabilistic deep learning approach for fluorescent particle tracking that quantifies uncertainty.
• To develop a method that does not require manual parameter tuning or prior knowledge of noise statistics.

Main Methods:

• A recurrent neural network mimicking Bayesian filtering was developed for particle tracking.
• The approach incorporates both aleatoric and epistemic uncertainty quantification.
• A novel neural network was introduced for joint correspondence finding and missing detection probability estimation.

Main Results:

• The method demonstrated state-of-the-art performance on synthetic and real 2D/3D fluorescence microscopy data.
• It successfully tracked intracellular structures, virus particles, and chromatin dynamics.
• The approach provides reliability information for computed trajectories.

Conclusions:

• The proposed probabilistic deep learning method offers an accurate and robust solution for fluorescent particle tracking.
• It significantly advances the quantification of dynamic biological processes in microscopy.
• The method reduces the need for manual intervention and prior data knowledge.