An essential first step towards understanding intracellular dynamic processes using live-cell time-lapse microscopy imaging is to extract accurate trajectories of all relevant particles in the images. One of the key aspects of this task is to make accurate associations between particle detections across time frames. State-of-the-art methods for this purpose often have many user parameters, sometimes even without a clear biophysical meaning, and/or they use explicit particle motion models that may not reflect reality, making them unfavorable for non-expert users or specific applications. Here we present a novel approach to data association for particle tracking applications based on deep neural networks. Specifically, we propose a recurrent neural network that learns particle behavior from the data, and based on this it determines how to best extend trajectories from one frame to the next. The results of preliminary experiments indicate that our method performs comparable to state-of-the-art data association methods for particle tracking, but with the advantage that it does not require users to tune any parameters.

Additional Metadata
Keywords Data association, Deep learning, Multi-task learning, Neural networks, Particle tracking
Persistent URL dx.doi.org/10.1109/ISBI.2018.8363615, hdl.handle.net/1765/108749
Conference 15th IEEE International Symposium on Biomedical Imaging, ISBI 2018
Rights no subscription
Citation
Yao, Y, Smal, I, & Meijering, E. (2018). Deep neural networks for data association in particle tracking. In Proceedings - International Symposium on Biomedical Imaging (pp. 458–461). doi:10.1109/ISBI.2018.8363615