Discovering Hidden Stomach Fibroblasts with Dana-Farber and BioTuring

Recovering hidden fibroblasts by letting each cell’s transcripts draw its boundary

Fibroblasts are among the hardest cells to segment in spatial data. They are long, thin, and threaded between other cells, while most segmentation models assume compact, round shapes. In mouse stomach corpus tissue, standard segmentation missed nearly half the fibroblasts, along with entire populations such as myofibroblasts and mesothelial cells.

Scientists at BioTuring tackled this by changing the starting point. Instead of guessing each cell’s shape from the image, they begin at the nucleus and let the cell’s own transcripts decide where it ends. The result: nearly 5,000 fibroblasts recovered, two missing populations restored, and a tissue map clean enough to resolve nine spatial niches along the gland.

Figure 1. Fibroblast segmentation in mouse stomach corpus Xenium data, following fibroblast-specific transcripts (dots) along their elongated shape.

We recently presented this problem together with Dr. Ke Li at Dana-Farber’s Spatial Transcriptomics Working Group, in a session titled Decoding the Glandular Stomach Microenvironment (October 21, 2026), hosted with 10x Genomics.

This work began with a question from Shivdasani Lab, Dana-Farber Cancer Institute, whose lab studies how stomach fibroblasts shape the epithelium around them. These fibroblasts send key signals, such as BMP near the top of the gland and WNT and R-spondin at its base, that guide epithelial differentiation. 

Why Fibroblasts Slip Through Segmentation

With a custom 100-gene Xenium panel, Cellpose segmented the stomach epithelium beautifully, cleanly separating the pit, isthmus, neck and chief zones. Fibroblasts were another story. Segmentation drew them as square or round cells, much like their epithelial neighbors, and missed many of them in the neck and chief zones entirely.

This matters because a boundary decides which transcripts belong to which cell. With boundaries that did not follow the cell, signals like BMP5 could not be reliably traced back to the fibroblast that produced them. Image-based models, trained mostly on compact cells, struggle with thin, irregular shapes whose stain signal is weak and easily cut short.

Figure 2. With default segmentation, Pdgfra+ fibroblasts (green) appear as compact, epithelial-like shapes near Bmp2 and Bmp5 transcripts, and are largely missing from the deeper zones. Source: Li et al., 2026, unpublished.

Our Approach: Growing Cells from the Nucleus Out

Our key insight: follow the signal, not the shape. A nucleus is easy to find reliably. Once we know where a cell starts, the transcripts around it can show how far it extends.

We first build a cell-type-specific marker list from a reference single-cell database; users will also be able to supply their own. Segmentation then runs in four steps:

1.     Group raw transcripts into small spatial bins.

2.     Seed one cell per nucleus from DAPI segmentation.

3.     Assign a cell type to each seed.

4.     Score neighboring bins for marker enrichment, accept or reject them, and repeat until no more bins pass the threshold.

Each cell grows outward from its nucleus to its full boundary, based on what it actually expresses. For fibroblasts, that means following fibroblast-specific transcripts into their true, elongated shape.

Figure 3. The transcript-guided segmentation pipeline: an offline marker-preparation stage feeds an online stage that expands each DAPI-seeded cell bin by bin.

Transcript-guided segmentation is still in development and not yet officially available in SpatialX. For other irregular cells, SpatialX already offers star-like segmentation (see our astrocyte work).

What the corrected map reveals

Better boundaries changed what the data could show. The transcript-guided approach identified 4,979 fibroblasts, compared with 2,506 from generic segmentation, nearly double. It also recovered 1,160 myofibroblasts and 249 mesothelial cells, two populations whose generic segmentation missed completely.

Figure 4. Cell type annotation with transcript-guided segmentation (left: 4,979 fibroblasts, plus myofibroblasts and mesothelial cells) vs. generic segmentation (right: 2,506 fibroblasts, no myofibroblasts or mesothelial cells).

With those cells in place, downstream analysis in SpatialX surfaced several exploratory hypotheses:

•      Spatial niches: neighborhood analysis at a 50 µm radius resolved nine niches that follow the gland’s architecture, from the pit surface down to the smooth muscle.

•      Fibroblast heterogeneity: the same cell type showed different programs by niche. Fibroblasts in the isthmus carried proliferation signatures, while those in the deep stromal core expressed structural matrix genes.

•      Immune neighbors: distance query showed macrophage density rising toward the isthmus niche boundary, while T cells were relatively more represented inside it.

Figure 5. Neighborhood analysis resolves nine spatial niches along the gland, each with its own cell type composition.

One audience question captured why boundaries matter here: how do you measure distance from a cell that touches dozens of neighbors? SpatialX measures from the nearest polygon edge, not the cell center, and that only works when the edge is right. All of these findings are hypotheses for future functional testing.

What this means for spatial biology

Segmentation is not just preprocessing. It is the first biological decision in a spatial study, determining which cell types exist in your data, which niches form, and how cells relate to one another. For stomach biology, accurate fibroblast boundaries open the way to studying how these cells support tissue homeostasis. The same transcript-guided idea can extend to other cells that don’t fit the round-cell mold.

Stuck on cells that don’t fit the circle? This project started with one lab’s segmentation problem, and we’d love to hear about yours.

Share your cell segmentation challenge and request a demo session tailored to your tissue.