Comparative Segmentation Analysis: MorphoGraphX and PlantSeg

To compare the segmentations created by MorphoGraphX alone with the ones using PlantSeg’s files as input, we first obtained a ground-truth segmentation using the MorphographX auto-segmentation pipeline as described in Strauss et al., 2019 (link) (Figure 10B) and manually fixed all segmentation errors using processes in MorphoGraphX. We then fed the confocal stacks to PlantSeg to compute wall predictions and 3D segmentations using the network trained on the ovule confocal data and the GASP method. Note that for samples with weaker cell wall signal we processed the raw input data in MorphoGraphX by adding a 2 µm thick layer of signal under the surface mesh and fed these to PlantSeg which tended to improve the PlantSeg output greatly. We then created surface segmentation using three methods: First, using the raw stack and the auto-segmentation pipeline in MorphoGraphX (method RawAutoSeg, Figure 10B, top). Second, using PlantSeg’s wall prediction values as input for the auto-segmentation process in MorphoGraphX (method PredAutoSeg, Figure 10B, left red arrow) and third, using PlantSeg’s fully segmented stack and projecting the resulting 3D labels onto the surface mesh using a custom process in MorphoGraphX (method Proj3D, Figure 10B, right red arrow).