A CNN-based cell tracking method for multi-slice intravital imaging data

Cell migration is one of the important criteria for determining effects on cells by inflammatory and/or chemical stimulation. Accurate detection of cells' movement through traditional methods, such as optical flow, is difficult because those cells' fluorescence intensity and shapes are similar to each other. Therefore, we adopt a tracking approach based on a convolutional neural network (CNN) using time-lapse multi-slice images observed with 2-photon excitation microscopy. Existing CNN-based cell tracking methods are often focusing on tracking targets in the 2-dimensional (2D) space since the costs for computation and annotation drastically increase in 3-dimensional (3D) settings. Those methods usually convert the 3D microscopic images to the 2D ones via maximum intensity projection (MIP). However, as MIP does not keep the depth information, it is difficult to track depth-directionally overlapping cells in MIP images accurately. To cope with the problem, we propose a novel CNN-based cell tracking method for multi-slice 3D images. Our method trains a CNN using MIP images annotated with the cell locations, similarly to the existing methods. Meanwhile, in the tracking phase, our method estimates not only each cell's location but its depth using multiple slices at different depths. Using our method, we track leukocyte migration in multi-slice time-lapse images observed with 2-photon excitation microscopy. As a result, we show that our method outperforms existing cell tracking methods including multi-domain network (MDNet) with MIP images.