The problem. In imaging assays you don’t get cells — you get a cloud of transcript coordinates. Assigning each molecule to the right cell is the load-bearing step, and doing it off nuclear stains alone throws away the molecular signal and misassigns transcripts at every boundary.

The idea. Baysor segments from the molecules themselves. It models the joint distribution of transcript positions and identities as a Markov random field, so a molecule’s neighbors and its expression type both inform which cell it belongs to. Nuclear or membrane stains, when available, enter as a prior rather than the sole truth. The output is cell boundaries that respect the actual RNA, not just the DAPI.

Why it matters. This is the step where a Xenium or MERSCOPE dataset silently goes wrong — over- or under-segmentation contaminates every downstream count, cluster, and cell-type call. For a facility, owning segmentation quality is non-negotiable, and Baysor is the reference point I’d benchmark any newer method against. It’s also the natural sequel to yesterday’s QC reading: QC tells you if segmentation failed, this tells you how to do it better.

Verdict. Foundational and still the default, though it’s compute-hungry and its priors need tuning per assay. Read it as the segmentation baseline; the newer graph-based methods are the next two entries, and this is what they’re trying to beat.