Fig. 3.
Evaluation of 3D cell segmentation quality on 3D multiplexed tissue images without manual annotations. (A) The top two principal components (PCs) of the segmentation quality metrics for 3DCellComposer, based on different 2D segmentation methods across three organs. The percentages next to PC denote the variance that each corresponding PC accounts for. These results are compared against Cellpose 2D to 3D and two direct 3D methods ACSS, and 3DCellSeg. Each method is represented by a unique color and distinct marker, while filled markers correspond to original images. Hollow marker trajectories illustrate varying levels of Gaussian noise perturbation, arranged in ascending order from low to high. The suffixes “mem” and “cyto” after DeepCell denote the cell membrane and cytoplasmic input channels, respectively. (B) A ranking of 3D segmentation methods by overall segmentation quality score. The green bars are 2D methods through 3DCellComposer whereas blue ones are other methods.