Figure 3.

Performance comparison of the models trained on the phase contrast microscopy dataset

(A) Average learning curves of U-Net and VGG19D-U-Net models trained on 10 frames per movie in leave-one-movie-out cross-validation. Solid lines are average training loss, and dotted lines are average validation loss.

(B) Average F1 scores of models trained on different numbers of frames per movie. Error bars: 95% confidence intervals of the mean.

(C) Training efficiency of models in terms of their model size, training time, and segmentation accuracy. The name of the model and number of parameters in italics are written on the bubble. The size of a bubble is proportional to the number of parameters in the model.

(D–F) Average F1, precision, and recall of models. For U-Net, suffix P denotes pretrained and no suffix P denotes non-pretrained model. Other models without suffix P are pretrained and have a U-Net decoder, same as the U-Net model. Suffix D denotes dropout layers added to the model. Significance was tested by the two-sided Wilcoxon signed-rank test. ns, p ≥ 0.05; ∗p < 0.05; ∗∗p < 0.0001. Error bars: 95% confidence intervals of the bootstrap mean. For (B–F), the number of evaluated frames is n = 202, which is roughly 40 frames from each phase contrast live cell movie.