A neural network model enables worm tracking in challenging conditions and increases signal-to-noise ratio in phenotypic screens
Fig 3
Tracking in challenging conditions.
(A) Multiple overlapping worms with high density of eggs. Inset images show higher magnification images of two sets of overlapping worms where all individuals are successfully skeletonized. (B) Examples of continuous tracks that preserve worm identity during and through collisions. (C) For highly curved worms that form long-lived coiling shapes, long gaps in the data can be present in Tierpsy-derived data. Here, there is a long gap with a high curvature that is recovered using DTC. (D) Improved skeletonization leads to longer tracks from DTC compared to Tierpsy. Note the log scale. The duration is longer for videos with 3 and 15 worms per well. (E) The number of skeletons/frame averaged over a video. DTC tracking produces numbers closer to the nominal number of worms per well. (F) The fraction of objects that is correctly tracked across frames compared to manually corrected trajectory data for videos with 3 and 15 worms per well.