Benchmarking cell segmentation methods for cardiac spatial transcriptomics using membrane staining as ground truth
H Tao, A LewisAbstract
Background/Introduction
Imaging-based spatial transcriptomics can resolve gene expression at subcellular level in intact tissue, but its value in human myocardium depends on accurate cell segmentation. Adult cardiomyocytes are large (100–150 µm), elongated, and often multinucleated; they bear no resemblance to the small round cells that segmentation algorithms were designed for. Default pipelines relying on nuclear staining provide poor membrane contrast in myocardium, and cardiomyocytes are routinely merged, fragmented, or missed. If the dominant cell type is systematically lost, any downstream analysis risks underrepresenting the very cells that define myocardial tissue.
Purpose
To benchmark multiple segmentation approaches in cardiac spatial transcriptomics using post-acquisition membrane staining as biologically defined ground truth, and to determine whether combining image-based and transcript-based methods improves cardiomyocyte recovery.
Methods
FFPE left ventricular tissue from patients with non-ischaemic cardiomyopathy (n=3, more incoming) was profiled using an imaging-based spatial transcriptomics platform (577 and 478 fields of view). Sections were stained post-acquisition with wheat germ agglutinin (WGA) and imaged by confocal microscopy. WGA images were registered to transcriptomics coordinates using DAPI as shared reference. Because WGA was applied to the same physical sections, the resulting membrane boundaries provide biological ground truth for evaluation. Seven segmentation algorithms are being benchmarked: five image-based (Voronoi, watershed, Cellpose, StarDist, Mesmer) and two transcript-based (Baysor, Proseg), each tested with and without membrane input. Alternative membrane markers including laminin are also under evaluation.
Results
In initial testing, WGA-guided Cellpose detected 69 cells per representative field of view versus 22 with the default DAPI-only pipeline, a 3.1-fold increase driven almost entirely by large cardiomyocytes that nuclear-only methods miss. WGA staining adds 2-4 hours of confocal time per slide. Systematic benchmarking of all seven methods, transcript-based refinement, and evaluation against morphometric and transcriptomic metrics are ongoing.
Conclusions
Standard segmentation substantially undercounts cardiomyocytes in cardiac spatial transcriptomics. WGA membrane staining provides both a practical route to better segmentation and a ground truth against which methods can be fairly compared. A systematic cardiac-specific benchmark is underway. Optimised cardiac-specific cell segmentation approaches are essential to realise the ability of spatial transcriptomics to define new biology of the human heart.Benchmarking workflow and proof-of-concept