Figure 5:

Effect of clustering on detection and interrater agreement. A. Stricter IOU thresholds reduce the number of anchor proposals generated by clustering but increase agreement. A threshold of 0.25 provides more anchor proposals with negligible difference in agreement from the 0.5 threshold. The shaded region indicates that by design, there are no anchor proposals with <2 clustered annotations. B. The clustering constraint prevents annotations from the same participant from being assigned to the same anchor, preserving participant intention when annotating overlapping nuclei. This results in better detection of overlapping nuclei during clustering (upper panel) and also affects the inferred P-truth for anchors (bottom panel). C. Interrater classification agreement among pathologists for tested clustering thresholds. D. Pairwise interrater classification agreement (Cohen κ) at 0.25 IOU threshold. **P < 0.01; ***P < 0.001.