Biology Atlas

Map / Concepts

concept · intermediate · partly written

Where the spatial tie is weakest: segmentation, registration and resolution

Imaging platforms leave 20% or more of transcripts unassigned by default segmentation, foreign-cell signal is common, and no micrometre registration error was found for cross-platform alignment.

Summary

Default segmentation leaves 20% or more of transcripts unassigned on each platform in the method authors' own benchmark. One lung preprint found 47.0 to 73.4% of some cell types with more than 5% foreign-marker signal. The only measured registration error found, 4.5 micrometres, is for mass-spectrometry imaging; none was found for cross-platform alignment. These are authors benchmarking their own tools. 123

Reported, not yet replicated

Transcripts to cells more

Proseg's authors report that default image-based segmentation leaves 20% or more of transcripts unassigned on each platform; Cellpose left about 26% in one MERSCOPE lung set. In a Xenium lung set, 37% of differential-expression calls were suspicious with the provided segmentation, 25% with Baysor and 12% with Proseg, where suspicious is the authors' own definition (gene pairs gaining over 50% co-expression after nuclear expansion). Baysor on CosMx lung predicted nearly 4 million cells, and about 27% of its cells were labelled ambiguous. The text could not be re-opened at the source check. 1

Reported, not yet replicated

Foreign signal inside cells more

In Xenium lung adenocarcinoma, 47.0 to 73.4% of B cells, T cells, fibroblasts and macrophages had over 5% of their transcriptome from other types' marker genes, against 3.0 to 10.3% in scRNA-seq; 0.2 to 2.9% had 25% or more. This is a preprint by the method's authors using five marker genes per type. They state that segmenters lack mechanisms to quantify the confidence of masks. 2

Reported, not yet replicated

Segmentation against human annotation more

On 31,384 manually annotated cells from 9 tissue-microarray cores on 3 platforms, precision in dense regions was 0.90 for CosMx and Xenium and 0.83 for MERSCOPE, and fell for sparse, then elongated cells. Between 2023 and 2024, total expression rose 4-fold on CosMx (authors credit an improved algorithm) and fell 8-fold on MERSCOPE (authors credit a changed baking time, not software). Read through a summarising tool and truncated. 4

Reported, not yet replicated

Protein imaging spillover more

In a simulation, classification AUROC was 0.94 to 0.98 at 1 micrometre; shrinking masks by one or two pixels cut total cell numbers by about 15 to 20% and 34 to 38%; and spillover cannot be corrected by mean aggregation. A preprint by the proposing group, with classifier-derived ground truth. 5

Reported, not yet replicated

Registration more

The OSTA book's registration chapter (read through a summarising tool) notes that automatic registration does not work well on images with very different orientation, scale and intensity, such as Xenium and Visium, and gives no error value. For mass-spectrometry imaging, off-tissue fiducials were not accurate enough; with manual control points and refinement the error was 4.5 plus or minus 1.45 micrometres against laser burn marks, under half of a 10 micrometre pixel, and the authors call precise co-registration the next bottleneck. A 2026 benchmark of 11 alignment methods includes a cross-platform scenario, but its figures could not be opened. No micrometre error for Xenium-to-Visium or H&E-to-Xenium was found. 36

Reported, not yet replicated

Resolution mismatch more

Visium spots of about 55 micrometres hold one to ten or more cells, and Visium HD 8 micrometre bins often hold transcripts from two to four cells. Section thickness differs by platform, so adjacent platforms measure different cells. 7

Reported, not yet replicated

Plain answer

In imaging-based spatial omics, a table row is only as true as the segmentation behind it, and published rates of transcripts left unassigned or assigned to a foreign cell are large; registration error between platforms is not reported in micrometres. 12

Reported, not yet replicated

Why it matters

Cell types, neighbour effects and ligand-receptor inference all inherit these errors, and the data format carries no flag for them. 2

Our inference

Real numbers

Unassigned transcripts 2026-11-18

20% or more by default segmentation on each platform; about 26% for Cellpose in one MERSCOPE lung set. 1

Reported, not yet replicated

Foreign-lineage signal 2026-11-18

47.0 to 73.4% of some cell types with over 5% foreign-marker signal in Xenium lung adenocarcinoma, against 3.0 to 10.3% in scRNA-seq. 2

Reported, not yet replicated

Mass-spectrometry imaging registration 2026-11-18

4.5 plus or minus 1.45 micrometres against laser burn marks. 3

Reported, not yet replicated

Still open here

Not yet written: Equation.

Linked from

Sources

  1. Jones DC, Elz AE, Hadadianpour A, Ryu H, et al.. Cell simulation as cell segmentation. Nat Methods (2025). paper; opened 2026-11-18 all records citing this
  2. Yang Y, DePasquale E, Adeleke D, Xie X, et al.. MisTIC: Missegmented Transcript Inference Correction for Improved Spatial Transcriptomics Analysis. bioRxiv (2025). preprint, not peer reviewed; opened 2026-11-18 all records citing this
  3. Ščupáková K, Dewez F, Walch AK, Heeren RMA, et al.. Morphometric Cell Classification for Single-Cell MALDI-Mass Spectrometry Imaging. Angew Chem Int Ed Engl (2020). paper; opened 2026-11-18 all records citing this
  4. Wang H, Huang R, Nelson J, Gao C, et al.. Systematic benchmarking of imaging spatial transcriptomics platforms in FFPE tissues. Nat Commun (2025). paper; opened 2026-11-18 all records citing this
  5. Gerber R, Griner J, Incicau D, Guglietta S, et al.. Spatially-smoothed quantification improves cell typing in imaging mass cytometry datasets. bioRxiv (2025). preprint, not peer reviewed; opened 2026-11-18 all records citing this
  6. OSTA book: coordinate registration chapter.. web, not peer reviewed; opened 2026-11-18 all records citing this
  7. Cervilla S, Grases D, Perez E, Real FX, et al.. A technical comparison of spatial transcriptomics platforms across six cancer types. Genome Biol (2026). paper; opened 2026-11-18 all records citing this
Where it sits on the map
History of this record (1)

This record has been changed 1 time, on the dates below. The note kept against each change is written for the people who maintain the atlas, in their own shorthand, so it is not shown here; it is in the record itself.

Status
partly written
Last updated
2026-11-18
Record
content/concepts/segmentation-and-registration-error-in-spatial-omics.json