Spatial profiling of transcriptional states in 3D provides valuable insights into cellular interactions in the context of tissue microenvironment. Accurate 3D cell segmentation is crucial for analyzing and creating cell-by-marker matrices from raw fluorescent staining signals. However, current 3D methods face challenges due to a lack of generalizability caused by limited training data and variations in image resolutions along both the XY-plane and Z-axis, unlike recent achievements in 2D segmentation.
To address these challenges, the Azizi and Blumberg groups collaborated to design CellStitch, a flexible and lightweight framework for 3D segmentation leveraging optimal transport. The study was published last month in BMC Bioinformatics. The method can be used to interpolate isotropic 3D cell morphology from highly anisotropic cell images.
Combining this with the state-of-the-art 2D segmentation approach, the team formulated the complete 3D reconstruction problem for cell correspondence along the 3rd dimension as a bipartite graph alignment and solved it with optimal transport. They demonstrated that CellStitch is robust against various image anisotropy without requiring extensive 3D training data. The method has shown significant performance improvements over competing methods on eight individual 3D plant microscopic datasets with diverse anisotropy and cell shapes.
“Our method, CellStitch, bridges the gap between the well-studied 2D segmentation and the increasing demand for 3D cellular reconstruction. The method is especially applicable on anisotropic images, which is common in many shallow-depth 3D microscopic tissue images,” explains Yining Liu, a PhD student in computer science (co-advised by Andrew J. Blumberg and Itsik Pe’er) and one of the study co-authors.
“Our interpolation findings also provide potential insights for experimental designs, where cytoplasm signals could be stained in less dense layers than RNA/protein channels given the technical difficulties for multiplexed imaging,” adds Yinuo Jin, a PhD student in biomedical engineering (Elham Azizi’s lab) and co-author of the study.
In the next phase, the team aims to enhance CellStitch by refining it with optical flows utilizing signals from additional projections. This adjustment is intended to reduce ambiguity in the stitching process. Additionally, they are eager to extend the application of CellStitch to in situ 3D microscopic datasets, allowing for the exploration of the local microenvironment in both normal and diseased tissues.
