Stereoscopic cell tracking for evaluating cell motility and mobility validated by deep learning

Abstract

Extracting dynamic features of a cell plays important role in understanding cell response to internal or external perturbations, which can be both a painful and imprecise task as one makes it manually in ocular way. Instead of using complex methods, we introduce a simple approach that uses disparity maps for segmentation by means of sequential frame couples. In our approach, disparity maps provide three-dimensional clues that can be used for cell segmentation. One of the contributions of this work is to generate pseudo 3D cell database using cell video frame couples. In addition, the optical flow method is performed to understand the cell behaviour and local dynamic movements. A mask regional convolutional neural network (Mask R-CNN) approach that requires manual segmented dataset and long training time is used for comparison. Obtained disparity-based segmentation and optical flow data are blended to easily analyse and evaluate the cell motility and mobility. In order to validate the segmentation results, Jaccard similarity index method is applied. Consequently, we succeed in dynamic segmentation-based tracking for understanding the cell behaviour without video enhancement or preprocessing steps, such as colour adjustment, filtering, thresholding.