Fig. 5

Model Training. A shows the training and testing losses of each of the models over time. We see that the models similarly converged over time, showing there were no training folds significantly different from others for training. B shows the mean Dice and Jaccard scores for the models tested on the holdout test set for the rod and ring mask recovery. C A representative example outputs along with the relative labels for the rods and ring masks. D Shows the per-pixel AUC curve for each of the masks. E and F show the correlation of the derived measurements from the model segmentations versus the segmentation provided by an expert annotator. A strong correlation here shows not only an accurate segmentation, but also a biologically helpful model that can be used for downstream analysis. Scale bar represents 30 μm