RNA assignment and segmentation geometry in Visium HD profiles of fibrostenotic Crohn's disease

Abstract

Retrospective reuse of spatial transcriptomics often starts from a deposited bin-level Seurat object rather than the inputs needed to revisit upstream molecular processing. We tested whether image segmentation could support a defensible cell-like reconstruction under this constraint in two complete H&E-stained intestinal sections from a Visium HD study of Crohn's disease-associated fibrostenosis (GSE302569). The available data comprised 18,085 genes, 988,942 original 8 {micro}m bins and 179,596,575 UMIs. We evaluated 20 segmentation configurations in the preserved image coordinate system using a common bin-centre assignment rule, without restarting Space Ranger or introducing higher-resolution expression. RNA assignment ranged from 0% to 74.68% in CD1 and 0% to 64.05% in CD2. Expansion increased recovery but did not eliminate merged regions, straight boundaries or within-bin ambiguity; approximately 70-73% of InstanSeg brightfield nuclear regions received no assigned RNA. All 38 non-empty configuration-slice combinations underwent native full-mode RCTD against the same nine-group scIBD reference. Mean slice-level RNA recovery was negatively associated with primary-component dominance across 19 configurations (descriptive Spearman {rho} = -0.77). Neither dominant-call agreement nor concentrated RCTD support validates boundaries or cell identity. The observed limitations in RNA recovery, region geometry, and compositional support did not establish any tested configuration as a reliable single-cell reconstruction from these inputs. The findings identify limitations of retrospective single-cell reconstruction from the available binned inputs, not the impossibility of cell-resolved analysis from suitable original data.