Segmenting budding yeast mother-bud pairs with SAM
Cell segmentation — defining cell outlines in microscopy images — is essential for quantitative image analysis. Budding yeast is a hard case: asymmetric cell division and mother-bud morphology cause dividing cells to be misidentified as two separate cells, which breaks downstream analysis. It is tempting to treat a general cell segmenter, or a vanilla Segment Anything Model (SAM), as good enough and to move on. This post stays inside one published study and asks a narrower question: when the error that matters is a split mother-bud pair, what does a yeast-specific SAM adaptation actually report?
The paper is Zhao et al., YeastSAM: A deep learning model for accurate segmentation of yeast cells, Molecular Biology of the Cell 37(6):mr5, 2026 (DOI 10.1091/mbc.E25-09-0454). The code is at https://github.com/YonghaoZhao722/YeastSAM, and the documentation is at https://yeastsamdoc.readthedocs.io/.
The setting, not a slogan
YeastSAM is a deep learning segmentation framework for budding yeast, built by adapting SAM. The point of the adaptation is mother-bud morphology: dividing cells should be kept as correct instances rather than split into two. Well-separated, non-dividing cells may not show the same error, so the practical gain is largest on dividing populations.
What YeastSAM is
The Crossref / journal abstract is direct: the authors addressed the mother-bud problem by adapting SAM to construct YeastSAM, a framework optimized for budding yeast. The open project is also a toolbox around that model. The GitHub repository and Read the Docs manual describe a GUI launcher and tools for generating masks, image registration, outline conversion, and mother-bud pair separation, written as a step-by-step guide for yeast segmentation and FISH-Quant integration. The documented tools support image series.
In the documented workflow, masks are generated from DIC images, typically in napari with custom YeastSAM weights (vit_b_yeast.pt, distributed on GitHub Releases). Those weights can also be used with BAND, the command-line interface, and QuPath. The docs recommend loading them rather than falling back to a default µSAM model. The repository acknowledges Segment Anything for Microscopy (µSAM; Archit et al., Nature Methods, 2025) as the inspiration; the published abstract describes YeastSAM as a SAM adaptation, and the open implementation is used with that microscopy SAM stack.
Mask2Outline converts masks into FISH-Quant-compatible outlines. A Mask Editor supports manual edits or automatic mother-bud separation with a documented CNN and U-Net module (models on GitHub Releases). Optional Shift Analyzer and Registration tools correct offset between DIC-derived masks and smFISH images. A training and validation dataset of DIC images and masks is on Zenodo (doi:10.5281/zenodo.17204942). The paper frames the tool as accessible to researchers with minimal programming experience; a macOS installer is provided in the releases.
YeastSAM is the segmentation component, not the RNA or organelle analysis, and it is not a general segmenter for non-yeast tissues.
What the abstract reports
YeastSAM achieved more than three-fold higher accuracy in segmenting dividing cells compared with existing methods. The project page does not invent additional unpublished metrics or per-dataset tables, and this post follows that boundary.
A bounded claim that stays faithful to the site and the abstract is therefore:
YeastSAM, built by adapting the Segment Anything Model (SAM) to budding yeast, achieved more than three-fold higher accuracy in segmenting dividing cells compared with existing methods, so dividing mother-bud cells are kept as correct instances rather than split into two cells.
That sentence is intentionally scoped. It names the SAM adaptation, the organism, and the dividing-cell comparison as reported. It does not say that the same factor will appear on every microscope or stain, or that well-separated G1 cells improved by the same amount. Exact dataset splits and baselines are in the full paper; “more than three-fold” is not a guarantee on a new setup without re-evaluation.
If you cite only “>3× more accurate,” please keep “dividing cells” and “compared with existing methods in the reported evaluation” in the same sentence.
The abstract also states a downstream use, not just a segmentation score. When combined with single-molecule RNA imaging and organelle imaging, YeastSAM can be incorporated into a computational pipeline to build cellular maps, enabling quantitative analysis of the spatial regulation of gene expression. YeastSAM itself is the segmentation component; map-building and biological interpretation depend on the user’s downstream tools, imaging quality, and analysis code.
What has to be true for this to be relevant
Three practical conditions sit behind the result.
First, the images should be budding yeast, and the error that matters should be instance identity of dividing mother-bud pairs. If you only need a coarse foreground mask, a yeast-specific SAM adaptation is not what the paper is arguing for.
Second, follow the documented install and weights. Hardware, GPU, and image-format requirements are in the docs, not in this post. Skipping the custom yeast weights is not the evaluation the paper is about.
Third, if you care about smFISH or organelle channels, plan for registration. The documented pipeline generates masks from DIC and then, if needed, corrects offset before converting to FISH-Quant outlines. Segmentation accuracy on DIC does not automatically give you aligned RNA spots.
When not to use it
Skip YeastSAM if your images are not budding yeast, if you only need coarse foreground masks without instance identity, or if a validated in-house model already handles mother-bud pairs on your setup. The paper’s claims focus on budding yeast dividing-cell accuracy and accessibility, not universal microscopy segmentation across tissues.
Domain shift across microscopes and stains can reduce the reported “more than three-fold” gain until the method is fine-tuned or re-benchmarked on local data. Downstream cellular-map analyses depend on RNA or organelle imaging quality and on the user’s own pipeline.
General tools — for example Cellpose-style models or vanilla SAM — are broader. The project page lists them as complementary context, not as a scored leaderboard, and this post does not invent comparison numbers for them.
A cautious way to use the work
If you image budding yeast and downstream quantification fails when a mother and bud are split into two cells, the design in the paper is worth trying: run the documented YeastSAM workflow, inspect dividing cells specifically, and only then pass masks into RNA or organelle analysis. Compare against whatever you currently use on the same dividing-cell images, rather than quoting the published factor as if it had already been measured on your microscope. The implementation and docs are in the YeastSAM repository and at yeastsamdoc.readthedocs.io.
If your cells are not budding yeast, this paper does not give you a pretrained generalist. If a local model already keeps mother-bud pairs intact, it does not argue that you must switch. If you do have the yeast morphology problem, the evidence is exactly as wide as the study: a SAM adaptation optimized for budding yeast, more than three-fold higher dividing-cell accuracy versus existing methods in the reported evaluation, and a documented path into cellular maps with single-molecule RNA and organelle imaging.
Reporting the experiment so others can judge it
If you benchmark a similar yeast segmenter, the comparison that matters is dividing-cell instance quality, not only whole-field scores on easy, well-separated cells. Write down the microscope, the imaging modality (the documented mask workflow is DIC), whether custom yeast weights were used, and how mother-bud correctness was scored. If you also report RNA or organelle maps, say how masks were registered to those channels.
Do not collapse the paper’s claim into “YeastSAM is three times better at cell segmentation.” The published factor is about dividing cells, relative to the paper’s existing-method comparison.
A note on scope for practitioners
People may want to cite this paper for a very broad sentence: foundation segmentation models can be adapted to difficult cell morphologies. That sentence is downstream of the result, but it is broader than the evidence. The evidence is: budding yeast, mother-bud errors, a SAM adaptation, more than three-fold higher accuracy on dividing cells versus existing methods, and an accessible workflow that can sit in front of single-molecule RNA and organelle imaging. If your use case matches that structure, the paper is directly on point. If it does not, treat the citation as a related-work pointer and look for evidence in your own domain.
The journal version, with full details and limitations, is DOI 10.1091/mbc.E25-09-0454.
Top comments (0)