Figure 2

Figure 2

Model comparison and representative slices. [Note (a) shows the accuracies of different cell segmentation models for the HMS dataset. 3DCellSeg achieves the second best accuracy in ARE, VOI_split, and VOI_merge, and achieves the best accuracy in Avg JI, JI > 70%, and JI > 50% (the plots for DSC-related metrics are of high similarity to JI-related metrics). (b) and (c) show representative slices of different model segments. ACME tends to under-segment (see the dark green region which mis-classifies different cells as one cell) while U-Net + SWS tends to over-segment (see the over-segmented small cells in the central region). PanopticFCN, Mask R-CNN FPN, and Mask R-CNN C4 are accurate on the HMS dataset but they are severely under-segment on the ATAS dataset. The cellular images in (b) and (c) were generated by Python Matplotlib (https://matplotlib.org) using the HMS and ATAS⁴⁹ datasets].