DeepKymoTracker: A tool for accurate construction of cell lineage trees for highly motile cells

Abstract

AbstractTime-lapse microscopy has long been used to record cell lineage trees. Successful construction of a lineage tree requires tracking and preserving the identity of multiple cells across many images. If a single cell is misidentified the identity of all its progeny will be corrupted and inferences about heritability may be incorrect. Successfully avoiding such identity errors is challenging, however, when studying cells such as T lymphocytes. These are highly motile and readily change shape from one image to the next. To address this problem, we developed DeepKymoTracker, a pipeline for combined tracking and segmentation. Central to DeepKymoTracker is the use of a seed, a marker which transmits information about cell position and identity between sets of images during tracking, as well as between tracking and segmentation steps. The seed allows a 3D convolutional neural network (CNN) to detect and associate cells across several consecutive images in an integrated way, reducing the risk of a single poor image corrupting cell identity. DeepKymoTracker was trained extensively on synthetic and experimental T lymphocyte images. It was benchmarked against five publicly available, automatic cell-analysis tools and outperformed them in almost all respects. The software is written in pure Python and is freely available. We suggest this tool is particularly suited to the tracking of cells in suspension, whose fast motion makes lineage assembly particularly difficult.Author SummaryCell lineage trees provide a striking visual representation of cellular decision-making and fate determination in a proliferating clone. Traditionally, these trees have been recorded using time-lapse microscopy movies of a dividing clone. Extracting lineage trees of T lymphocytes from microscopy movies is particularly challenging, however, because the cells are highly motile, readily change shape, and are easily damaged by excessive illumination and frame rates. Here we present a deep-learning approach to extracting cell lineage trees from movies of proliferating T cells. Using 3D convolutional neural networks for tracking and separate networks for segmentation we were able to reduce misidentification errors and record cell lineage trees with high fidelity. In benchmark tests, our algorithm was found to outperform all other state-of-the-art algorithms in both tracking and segmentation.