Fig. 1

Overview of CellStitch framework. a CellStitch consists of a 3-step pipeline to reconstruct 3D cells: backbone 2D segmentation along Z-axis direction, stitching module, and an optional interpolation module. b Stitching pipeline: CellStitch first computes the source and target distributions based on the cell masses on given adjacent layers, and a cost matrix based on the pairwise overlap. The solved optimal transport plan is then used to deduce the optimal correspondence of cells across the adjacent layers. Finally, it reassigns instance labels of cell slices to enforce labeling consistency across all layers and creates new labels when new cells emerge based on the segmentations masks on the other two projections. c Interpolation pipeline: After finding optimal instance stitching, CellStitch leverages pixel-wise boundary matching followed by morphology interpolation to predict the internal layers between adjacent slices