On The Preparation Of Training Data For 3D Tracking Of Position And Orientation Of Microswimmers Using Defocusing And Deep Learning

Abstract

The use of image analysis algorithms based on deep convolutional neural networks (CNNs) is opening new possibilities in experimental techniques for fluid dynamics. In the field of microscopic 3D particle tracking based on defocusing, it has been shown on synthetic images that CNNs can be used to determine not only the depth position but also the orientation of non-spherical microscopic particles or specimens. However, the translation of such approaches to experimental setups is challenging, especially due to the difficulty of obtaining reliable labeled images to train the neural networks. In this work, we propose a method to obtain labeled images of microswimmers from conventional microscopy images, exploiting the properties of the swimming motion at low Reynolds numbers, where inertial effects are negligible. This method allows us to obtain ground truth values of orientation and depth position of a microorganism swimming at a regular pace that can be used as training data for CNNs. The proposed methodology was used to obtain labeled images from experimental videos of the micro-organism Euplotes Vannus. The labeled images were used to train a VGG16 network, providing an estimated uncertainty in the determination of the orientation angles psi, theta, phi, and the depth position Z of 2.4%, 1.4%, 2.3%, and 3.5%, respectively.