Recurrent Neural Networks for Lagrangian Tracking of Bacteria

Abstract

Research in microbiology is crucial for development of antibiotics, vaccines and other medicines that cure diseases and prevent spread of viral infections. One method of studying microorganisms is Lagrangian tracking, where the movement of single microorganisms, such as bacteria, is tracked over long periods of time, which is important when studying for example chemotaxis. Lagrangian tracking has previously been implemented using deep learning, showing promising results. However, the model, a convolutional neural network (CNN), struggled to handle overlapping bacteria, which resulted in failure of entire experiments when the model switched which bacterium was currently being tracked. This thesis aimed to create a model for Lagrangian tracking that could accurately track over long periods of time as well as handle overlapping bacteria. The method included simulation of fluorescence microscopic data as well as design, training and evaluation of recurrent neural networks (RNNs) using the simulated data. The results showed that the RNNs gave lower error distributions and were able to handle overlapping bacteria better compared to the CNNs implemented for benchmarking. An analysis of the importance of features of the bacteria for tracking indicated that the tracking was harder when surrounding bacteria were close to the focal plane or had higher intensity compared to the bacterium that was currently tracked. Although testing the RNNs in an experimental setup remains, the results suggest that replacing a CNN with an RNN can improve the accuracy of the Lagrangian tracking and to greater extent avoid losing the bacterium during an overlap. In turn, improving the accuracy of Lagrangian tracking contributes to the possibility of tracking single microorganisms over long periods of time and gain more knowledge about for example chemotaxis.