A neural network model enables worm tracking in challenging conditions and increases signal-to-noise ratio in phenotypic screens
Fig 2
(A) Histograms of root mean square deviation between held-out skeletons and predicted skeletons for each tested model. The vertical dashed line shows the RMSD between three manual annotators. Skeletons above the histogram are examples that illustrate the corresponding RMSD visually. (B) Bar chart showing the number of cases in the held-out test data where a model fails to make a prediction (e.g., Tierpsy fails on coiled worms or a neural network model does not identify any worms above a confidence threshold). (C) Examples of the kinds of errors each model makes. PAF and Omnipose often over-segment worms. DTC fails when worms are fully coiled in circles or in tight parallel contact for an extended time. (D) Computation time per input frame for the different models as a function of worm number/well. Each camera records 16 wells in a 96-well plate so these correspond to 48 and 240 worms per video. Tierpsy only uses CPU computation while Omnipose uses GPU and CPU because we use Tierpsy’s skeletonization algorithm to convert segmented regions to skeletons.