CellPoseSAM: SAM-Based Cell Segmentation

• CellPoseSAM is a microscopy-specific instance segmentation model that reuses SAM’s image encoder and trains a domain-specific decoder for cell-related instance prediction.
• It employs CellPose-style postprocessing to automatically convert predictions into final masks without relying on explicit prompting, enhancing automation in segmentation.
• Evaluated on 36 diverse datasets, CellPoseSAM demonstrates competitive performance in label-free, fluorescence, and histopathology tasks, establishing it as a strong baseline.

Searching arXiv for CellPoseSAM and closely related SAM-based cell segmentation papers to ground the article in current literature. I’m checking whether the environment exposes an arXiv search utility. CellPoseSAM is a microscopy foundation model for instance segmentation built by taking a SAM image encoder and training a new decoder to predict cell-related instance information, then converting those predictions into final instance masks with non-differentiable post-processing. In a comparative taxonomy of SAM adaptation strategies, it corresponds mainly to using SAM’s encoder as a backbone and training a domain-specific decoder for automatic segmentation. Comparative studies place it among the strongest microscopy-specific baselines across cell and nucleus segmentation, especially in settings where general-purpose SAM variants are not yet competitive with domain-adapted models (Archit et al., 18 Mar 2026).

1. Model identity and place within SAM-based microscopy segmentation

CellPoseSAM is most clearly defined in the comparative study "Revisiting foundation models for cell instance segmentation" (Archit et al., 18 Mar 2026). There it is treated as a specific microscopy foundation model rather than as a generic label for any SAM-based cell segmenter. Its defining configuration is a SAM image encoder, a newly trained decoder, and CellPose-style postprocessing for instance extraction.

Within the same taxonomy, CellPoseSAM is contrasted with three other design patterns. CellSAM is characterized as encoder + box detector + SAM promptable mask decoder. $\mu$SAM is characterized as full SAM fine-tuning + custom decoder + watershed-based instance extraction. APG is described as a new strategy built on $\mu$SAM that uses model predictions to generate prompts automatically and then applies SAM-style prompt decoding plus NMS. This taxonomy is important because it separates automatic decoder-driven segmentation from detector-prompt pipelines and from prompt-generation pipelines, even though all of them are descendants of the broader SAM adaptation program for microscopy (Archit et al., 18 Mar 2026).

A recurring theme in the literature is that CellPoseSAM occupies the automatic-segmentation side of this design space. Unlike systems that depend on explicit box prompting or user interaction at inference, it is framed as a model that predicts instance-relevant information directly and then resolves those predictions into masks through a downstream algorithmic stage. This makes it particularly salient in benchmark comparisons that emphasize large-scale, test-split evaluation across heterogeneous microscopy domains.

2. Architectural logic and Cellpose lineage

The comparative literature describes CellPoseSAM at a relatively high level: a SAM image encoder is reused, a new decoder is trained, and final masks are obtained through CellPose-style postprocessing (Archit et al., 18 Mar 2026). The exact decoder internals are not specified in the provided material, but the architectural logic is explicit: pretrained foundation features from SAM are retained, while the segmentation-specific inductive bias is shifted into a domain decoder and the postprocessing stage.

The reference point for the “CellPose-style” descriptor is the Cellpose family of methods. In the 3D Cellpose extension, Cellpose is described as representing each cell instance via a vector field or flow field that points toward the cell center, together with a foreground map; segmentation is then reconstructed by iteratively following the flow field and grouping points or voxels that converge to the same center sink (Eschweiler et al., 2021). Because CellPoseSAM is described only as using “CellPose-style postprocessing,” this suggests a lineage to flow-based or center-seeking instance reconstruction, but the exact reconstruction variables used by CellPoseSAM are not specified in the provided data.

This architectural positioning distinguishes CellPoseSAM from prompt-centric SAM adaptations. CellSAM relies on an automated detection module to generate bounding-box prompts for SAM’s prompt encoder and mask decoder (Israel et al., 2023). SAC similarly emphasizes prompt quality, introducing an auto-prompt generator that produces positive and negative point prompts, alongside LoRA in SAM attention layers and a trainable mask decoder (Na et al., 2024). By contrast, CellPoseSAM is grouped with methods that primarily convert encoder features into automatic instance predictions and then resolve instances with postprocessing rather than with prompt-conditioned mask decoding as the central mechanism.

3. Benchmark protocol and quantitative standing

The most detailed quantitative characterization of CellPoseSAM comes from an evaluation over 36 datasets spanning four modalities or tasks: cell segmentation in label-free microscopy, cell segmentation in fluorescence microscopy, nucleus segmentation in fluorescence microscopy, and nucleus segmentation in histopathology. The study uses the test split for all datasets, and for volumetric or time-lapse datasets evaluation is performed slice-wise in 2D for efficiency. The primary metric is mean segmentation accuracy, defined as $$\text{Mean Segmentation Accuracy} = \frac{1}{|\text{\# thresholds}|} \sum_{t} \frac{TP(t)}{TP(t) + FP(t) + FN(t)}\,,$$ with thresholds $$t \in [0.5, 0.55, 0.6, 0.65, 0.7, 0.75, 0.8, 0.85, 0.9, 0.95].$$ This metric averages matched-instance quality over multiple IoU thresholds and is therefore stricter than a single-threshold overlap score (Archit et al., 18 Mar 2026).

On modality-averaged performance, CellPoseSAM is reported as follows: 0.544 on label-free cell segmentation, 0.363 on fluorescence cell segmentation, 0.483 on fluorescence nucleus segmentation, and 0.418 on histopathology nucleus segmentation. In the same benchmark, these values make it the best method on label-free cell segmentation, the best method on fluorescence cell segmentation, and the best method on histopathology nucleus segmentation, while APG slightly exceeds it on fluorescence nucleus segmentation with 0.518. The study therefore identifies CellPoseSAM as the strongest single baseline overall, while also concluding that APG is competitive with it in several settings (Archit et al., 18 Mar 2026).

The paper further states that CellPoseSAM is among the top three methods for all four modalities. Per-dataset highlights emphasize strong performance on several label-free datasets, including OmniPose, DeepBacs, and TOIAM, with a score of 0.898 on TOIAM. It is also reported as strong on TissueNet, CellPose, HPA, and CellBinDB for fluorescence cell segmentation; on DSB, U20S, and IFNuclei for fluorescence nucleus segmentation; and on LynSec, MoNuSeg, PUMA, TNBC, CryoNuSeg, and CytoDark0 for histopathology nucleus segmentation. In the same study, CellPoseSAM consistently outperforms SAM, SAM3, and CellSAM on aggregate, while SAM2 is not included in the final benchmark because it was found inferior to SAM in this setting and failed to segment anything on several datasets in initial experiments (Archit et al., 18 Mar 2026).

4. Histopathology, nucleus segmentation, and the broader pathology adaptation trend

Although CellPoseSAM is framed as a microscopy foundation model across multiple modalities, its role in histopathology is especially visible through comparisons with pathology-specific SAM adaptations. PathoSAM is described as conceptually close to the broader trend of foundation-model-based cell or nucleus segmentation, similar to CellPoseSAM-style efforts, in that both aim to move beyond task-specific segmentation toward more general, adaptable nucleus or cell segmentation. The key distinction drawn in that work is that PathoSAM is explicitly built by adapting SAM to pathology, with interactive prompting as a first-class capability, whereas CellPose-style methods are typically specialized segmentation systems optimized for microscopy morphology and often rely on image-specific training or heuristics (Griebel et al., 1 Feb 2025).

In that comparison, PathoSAM’s reported strengths relative to CellPose-like methods are promptable interactive segmentation, strong cross-dataset generalization, compatibility with annotation workflows, and one general model covering multiple nucleus domains. Relative to CellPoseSAM-like ideas, PathoSAM is described as placing stronger emphasis on interactive correction, AIS for automatic instance segmentation, and generalization across stains, tissues, and imaging conditions. This is a useful contrast because it clarifies that CellPoseSAM occupies the automatic-decoder branch of the field, whereas some later pathology models pursue a more explicit unification of automatic and interactive segmentation (Griebel et al., 1 Feb 2025).

The pathology literature also sharpens the distinction between instance segmentation and semantic nucleus classification. PathoSAM is reported as the new state of the art for automatic and interactive nucleus instance segmentation in histopathology, but not yet state of the art for semantic nucleus segmentation, where CellViT remains better. A plausible implication is that CellPoseSAM’s strong standing in instance-centric benchmarks does not automatically imply superiority in fine-grained semantic settings, since the field increasingly separates automatic instance delineation from downstream class-aware segmentation and classification (Griebel et al., 1 Feb 2025).

5. Related models and the ambiguity of “CellPoseSAM-style”

One source of terminological ambiguity is that several papers use “CellPoseSAM-style” as a comparative label for a family of SAM-adapted cell or nucleus segmentation approaches, even when they are not describing the specific benchmarked model CellPoseSAM. CellSeg1, for example, is explicitly described as “best understood as a CellPoseSAM-style, low-shot universal cell segmentation method”: it starts from SAM, fine-tunes with LoRA, uses only point prompts, and performs inference by sampling a $32 \times 32$ grid of points in SAM’s “everything” mode, followed by optimized NMS. Its headline result is an average $mAP_{0.5} = 0.81$ across 19 datasets when trained on one image, which makes it a low-shot adaptation strategy rather than a direct reimplementation of CellPoseSAM (Zhou et al., 2024).

SAC is another neighboring system that is said to be “very much in the same family” as a CellPoseSAM-style adaptation, but with a different emphasis. It inserts LoRA into SAM’s image encoder attention layers, trains a 4-layer U-Net auto-prompt generator, uses positive and negative point prompts, and fine-tunes the SAM mask decoder. On MoNuSeg, SAC 0-expert reports F1 84.11, IoU 72.61, and Dice 84.03; on DSB, it reports F1 93.48, IoU 87.32, and Dice 93.04. These comparisons show that some of the literature regards CellPoseSAM less as a single immutable architecture and more as a reference point for SAM-based biomedical instance segmentation with strong prompting or adaptation components (Na et al., 2024).

By contrast, the “molecular-empowered All-in-SAM” paper states more explicitly that it is adjacent to CellPoseSAM-style segmentation, but not a direct counterpart. Its novelty lies in a three-stage framework of molecular-empowered annotation, SAM adaptation with adapters, and Molecular-Oriented Corrective Learning, with paired molecular images used during training or annotation but PAS-only inference at test time. The paper also states that it does not present a direct comparison with CellPoseSAM and does not frame itself as a CellPoseSAM variant (Li et al., 21 Aug 2025).

This variability in usage means that “CellPoseSAM” can denote either a specific model class benchmarked as encoder + custom decoder + CellPose-style postprocessing, or, more loosely, a broader style of SAM-based cell segmentation that emphasizes automatic instance extraction and microscopy adaptation. The comparative review resolves this ambiguity by treating CellPoseSAM as a specific, high-performing baseline within a larger ecosystem of SAM adaptations (Archit et al., 18 Mar 2026).

6. Limitations, comparative lessons, and likely development trajectories

The comparative benchmark around CellPoseSAM draws several field-level conclusions. First, microscopy-specific adaptation still matters: general-purpose segmentation foundation models such as SAM, SAM2, and SAM3 are not enough for strong microscopy instance segmentation. Second, performance correlates strongly with the size and diversity of the domain-specific training set and with whether evaluation is in-domain or out-of-domain. Third, prompt strategy remains important even for automatic systems; the APG results show that automatic prompt generation can materially improve a SAM-style microscopy model without retraining (Archit et al., 18 Mar 2026).

These conclusions position CellPoseSAM as both a strong method and a useful reference architecture. It is one of the strongest current microscopy foundation models for cell and nucleus instance segmentation, yet it is also part of a moving target in which prompt generation, decoder design, and task specialization are still being reconfigured. The APG study, for example, argues that future cell instance segmentation models should integrate automatic prompt generation, strong domain-specific decoders, promptable SAM-style refinement, and possibly box prompts or iterative prompting. This suggests that the most consequential successors to CellPoseSAM may be hybrid systems that retain its strong automatic-segmentation backbone while importing some of the flexibility associated with promptable refinement (Archit et al., 18 Mar 2026).

Other neighboring results point in the same direction from different angles. CellSeg1 argues that rapid adaptation from one excellent image may be more practical than pursuing a wholly universal model, and demonstrates robust transfer with LoRA-adapted SAM trained from one carefully annotated image (Zhou et al., 2024). PathoSAM shows that a SAM-derived model can simultaneously support automatic instance segmentation and interactive correction in histopathology, with practical integration into QuPath and $\mu$SAM/napari (Griebel et al., 1 Feb 2025). Taken together, these results suggest that CellPoseSAM’s enduring significance lies not only in its benchmark standing but also in having established a particularly effective synthesis of SAM feature reuse and automatic, morphology-aware instance reconstruction.