Fig. 1: Method overview.
a, Current cell-tracking algorithms convert microscopy images into cell tracks without providing information on accuracy. Yet even single errors can greatly alter the biological interpretation of lineages (here, change in symmetry of divisions). Hence, extensive manual review is required and finally no assessment of statistical confidence can be provided. b, OrganoidTracker 2.0 outputs not only tracks but also associated error rate estimates, greatly aiding data interpretability and transparency. These error estimates also enable drastically reduced manual review or fully automated filtering to achieve high-confidence datasets. c, Method workflow, highlighting two new components (gray boxes): i, Generation of 3D confocal stacks of nuclear marker fluorescence. ii, Neural network detection of nuclear centers. iii, Neural network prediction of cell linking or division probabilities, based on image crops. iv, Constructing a graph representation of the tracking problem, based on predicted link and division probabilities. v, Determination of the globally optimal solution representing the most likely cell trajectories. vi, Estimating link error rates through systematic comparison with alternative tracking solutions. vii, Predicted cell tracks with error rate predictions for individual links.
