Fig. 5. Detailed computational comparison of all methods.

The evaluations were carried out on 17 test images of 1894 cycling wild-type cells not included in the YeaZ training set. a Each row shows an example test image, its ground-truth annotation (GT), and the result of Wood et al.⁹, YeastSpotter¹⁴, and YeaZ, respectively. b Quantification of segmentation performance of all methods. As is common in the computer vision literature, we call a predicted cell a true positive (TP), if its intersection over union (IoU) with the corresponding ground-truth (GT) cell is larger than or equal to 50%. Similarly, false positives (FPs) and false negatives (FNs) are defined as predictions that have no GT match and vice versa. As segmentation metric, we show the average accuracy ($\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}+\mathrm{FN}}$) and average intersection-over-union of true positives (IoU). Boxes show interquartile ranges (IQR), lines signify medians, and whiskers extend to 1.5 IQR. Scale bar: 5 μm.