Single camera estimation of microswimmer depth with a convolutional network.

Area of Science:

• Microbiology
• Biophysics
• Machine Learning

Background:

• Measuring protist 3D trajectories often requires complex experimental setups like orthogonal cameras.
• Existing machine learning methods for vertical position estimation need large, finely-spaced labeled datasets.

Purpose of the Study:

• To develop a simple, cost-effective method for generating labeled 3D trajectory data for protists.
• To train a machine learning model for estimating Paramecium's vertical position using standard bright-field microscopy.
• To apply this technique for studying Paramecium's thigmotaxis behavior.

Main Methods:

• A tilted slide setup was used to create a dataset of Paramecium images with labeled vertical positions from a single 5-minute movie.
• A simple convolutional neural network was trained on this dataset.
• The trained network was used to estimate Paramecium's vertical position from conventional bright-field images.

Main Results:

• A dataset of labeled Paramecium images was successfully generated using the tilted slide method.
• A convolutional network was trained to accurately estimate Paramecium's vertical position.
• The developed technique demonstrated sufficient accuracy for studying Paramecium's surface-following behavior (thigmotaxis).

Conclusions:

• The tilted slide method provides a simple and efficient way to create labeled datasets for protist tracking.
• Machine learning, using conventional bright-field microscopy, can accurately determine protist vertical position.
• This technique facilitates the study of behaviors like thigmotaxis in Paramecium.