Private cell quantification for microscopy images. Segmentation, per-cell measurements, assays, correction, and export on macOS, Windows, and the web.
Platforms · What it does · Complete feature list · Quick start · Limitations · Citing
CellCounter turns microscopy images into reviewable cell masks, measurements, size distributions, and assay results. It supports counts, size classes, marker-positive fractions, colocalization, confluence, wound closure, foci per cell, migration, and more through a focused graphical interface.
These analyses are usually done with an ImageJ macro or a CellProfiler pipeline. CellCounter does them through a graphical interface instead. It applies to any cell type a supported model can segment.
Image analysis is local by design. CellCounter does not require an account and does not upload microscopy images.
| Platform | Release | Availability |
|---|---|---|
| macOS | v1.0.13 | Native universal app for macOS 15 or later |
| Windows | v1.1.0 | Native x64 .exe and .msi installers for Windows 10 and 11 |
| Web | v0.2.1 preview | Installable PWA with local classical segmentation; learned models remain unavailable |
Note
The web preview keeps images in the browser and runs classical threshold/watershed segmentation, but its three learned segmentation models are not bundled yet. Use the macOS or Windows app when live model inference is required. Platform-specific details are available in the Windows guide and web guide. See the complete feature list for a release-by-release comparison.
Feature parity is partial. All three platforms now provide saved setup, preview, resumable processing, linked measurements, and mask review. Windows retains three CPU-only learned models; the PWA currently runs classical segmentation. The broader model catalog, microscopy workspace and several correction/training tools remain macOS-specific. See release status and remaining parity gaps for the current inventory and confirmation status of the two reported desktop issues.
A neutral, muted-blue workbench replaces the crowded interface. Preview and process with local classical segmentation, recover saved jobs, and review linked measurements or mask alternatives. A continuous, full-page 3D introduction follows the scroll from cell-like forms to an ordered field, with rotation, pause and reduced motion; library icons replace bespoke interface marks. See the web release notes.
The first-analysis “Sidecar scripts are not staged” error is fixed. Import into a saved setup, preview before processing, pause/resume jobs, and inspect linked measurements or saved mask versions. Large Review Queues use bounded pages and support Undo across pages. See the Windows release notes for training requirements, unsigned installer guidance and remaining platform gaps.
Large Review Queues use bounded contour caches and stable pages. Skip reaches every candidate, previews follow the current card, and saved edits plus undo remain reliable across pages. This addresses issue #9, reported after processing hundreds of images. See the release notes.
- Ordinary PNG, JPEG and BMP source previews use native decoding before Python startup; unsupported scientific layouts retain their source reader.
- Analysis setup starts with a valid image selection, keeps its preview visible while preparing a channel, and avoids redundant image reloads.
- First-use Cellpose weight downloads show progress and actionable network errors. Missing calibration metadata is explained separately from image or model failures.
See the hotfix release notes.
macOS v1.0.11 introduced the native inspect → preview → process → review workflow. Windows and the web have separate releases with the differences listed in the feature reference.
- The Models page shares background availability checks instead of repeating runtime imports on first navigation.
- Screen-aware keyboard commands appear in the menus and in Help → Keyboard Shortcuts (
⌘/), with standard text editing preserved. A new microscopy icon is included in the macOS app assets. - Import and inspect before installing a detector; choose a representative image and preview masks before a batch.
- Keep processing in a persistent queue, browse completed images, pause after the current image, resume after relaunch, and retry failures without repeating successful images. Interrupted work waits for you to resume.
- Select the same cells in an image, measurement table, or scatter plot; compare saved mask variants with linked pan/zoom and added, removed, and changed outlines.
- Start from cell counting, nuclei, marker positivity, or wound-closure tasks. Each result retains its original analysis settings; current calibration and review filters are shown separately.
- Train Cellpose 3.x from explicitly reviewed library masks, keeping specimen groups together across training, validation, and held-out evaluation. Activation requires a real checkpoint and a verified evaluation report.
- Reuse warm assay workers and bounded source/mask caches; draw visible masks using cached paths. Prepare the next source only within a conservative memory budget while model inference remains serial.
See the release notes and native workflow details for supported behavior and limits.
Cellpose provides its own graphical interface, and CellCounter does not replace it. It covers requirements the general-purpose tools do not provide directly:
| Installation without environment setup | The installer sets up the Python environment in the background. There is no conda, pip, or PATH configuration. |
| Size classes in addition to masks | Cells are binned into size categories in micrometres, using a pixel size read from the image metadata. |
| Assays in addition to counts | Marker-positive fraction, colocalization, confluence, wound closure, foci per cell, and others. |
| Local processing | Analysis remains on the device, helping laboratories keep sensitive image data under their own control. |
The macOS app offers the full set of model families summarized below. Windows v1.1.0 focuses on three validated families:
Cellpose-SAM (legacy cpsam_v2 ID), Cellpose cyto3, and StarDist fluorescence. The web preview exposes the same three-model catalog,
with learned inference clearly marked unavailable until validated browser weights are distributed.
| Model | Best for |
|---|---|
| Cellpose-SAM | Large or irregular cells — where cyto3 merges neighbours into one mask |
Cellpose cyto3 cyto2 nuclei |
General cytoplasm and nuclei |
| StarDist | Crowded, roughly convex nuclei |
| Omnipose | Bacteria and elongated / filamentous cells |
| Threshold + watershed | No weights, no download, no GPU — instant and deterministic |
| Your own model | Load a fine-tuned Cellpose checkpoint or a StarDist model directory |
Plus a second-opinion mode: run two detectors and review only the cells where they disagree. The macOS catalog also includes Cellpose-SAM v2, Cellpose-DINO ViT-L and ViT-B, multiple StarDist and Omnipose checkpoints, and Otsu, triangle, adaptive, and manual threshold variants. Models that are still being validated are marked as unavailable in the application rather than presented as runnable.
The native apps read JPEG, PNG, BMP, TIFF / OME-TIFF, and common microscope containers. macOS supports Zeiss
.czi, Nikon .nd2, Leica .lif, and Olympus .oif .oib .oir; Windows supports .czi, .nd2, .lif,
.oir, and .vsi. macOS offers max, sum, and mean Z-stack projections plus channel selection and naming;
Windows analyzes a chosen channel with maximum, mean, sum or middle-plane projection; its setup viewer still shows the imported display preview. The web preview supports JPEG, PNG,
WebP, BMP, TIFF, and OME-TIFF; proprietary microscope containers should be converted locally to OME-TIFF first.
Calibration can be read from compatible OME, ImageJ, TIFF, and microscope metadata or entered manually from a known pixel size or scale bar. Named calibration presets, configurable size bins, and analysis protocols make the same settings reusable across a study.
The macOS app includes a native, layer-based workspace for exploratory image analysis and presentation:
- Browse multidimensional time, Z, and channel axes without loading an entire sequence into memory.
- Open local OME-Zarr / OME-NGFF multiscale datasets and plates, with automatic resolution selection.
- Align image layers, create bounded-memory tile mosaics, and build or manually correct cell lineages.
- Paint class labels and export 16-bit training masks for use with a separate model-training workflow.
- Record repeatable local workflows for layer, axis, registration, and stitching operations.
- Save and reopen workspace projects, control layer visibility and opacity, and create keyframed GIF animations of axis, camera, visibility, and opacity changes.
Workspace projects and images stay on the Mac. The extension browser exposes only bundled, curated capabilities; it does not download or execute third-party plugin code.
|
Per cell
|
Per image
|
|
Fluorescence
|
Time series & morphology
|
- Library and batches — browse imported images, organize conditions, inspect batch summaries, and find exact duplicates.
- Focused review — triage low-confidence detections, attach per-image notes and review confidence, and keep correction history with undo and redo.
- Image inspection — switch channels and Z projections, inspect intensity histograms and calibrated line profiles, and apply include/exclude regions of interest.
- Quality control — review focus and illumination indicators, confidence and diameter distributions, drift, detector agreement, and fields ranked for attention.
- Detection refinement — adjust confidence and expected diameter, use preprocessing presets and background subtraction, split touching cells, and rerun detection without losing the source image.
Windows v1.1.0 supports local Cellpose cyto3 fine-tuning with a held-out test split, progress and cancellation,
versioned checkpoints, and explicit activation. macOS v1.0.12 trains compatible Cellpose 3.x models from reviewed
library masks, keeps independent specimen groups in separate train/validation/test partitions, and evaluates
the selected checkpoint on held-out images. Training requires at least three specimen groups and six epochs;
checkpoint activation checks the saved evaluation report and weight-file hash. Version history supports rollback.
The web preview does not train models; corrected label masks and ROI data can be exported for later workflows.
- Manual correction — add, delete, merge, split, resize, or trace a cell by hand. Corrections persist and the exported count is the corrected count.
- Prompt-guided correction — refine a mask with a point or box prompt. Compatible
micro_saminstallations reuse per-image embeddings; a native local fallback remains available. - Sequence correction — propagate selected masks, interpolate labels between frames, and compensate acquisition drift before tracking.
- Fast curation — review uncertain detections as cards or a tiled grid, compare reversible mask variants, and rank fields by heuristic error risk.
- Quality insights — cached confidence/diameter plots, drift trends, detector agreement, and preprocessing recommendations are recomputed only when a batch changes.
- Compare two conditions with a Mann-Whitney U test and effect size — read the limitations first.
- Score against ground truth — F1, precision, recall vs. your own hand counts.
- Export — PDF reports, annotated images, per-cell CSV, per-image summary CSV with one column per size bin, ImageJ ROI sets, GeoJSON (QuPath), and reproducibility metadata. macOS can also export ground-truth annotations and create a self-contained sample folder containing the source, overlay, tables, ROIs, Markdown and PDF reports.
- Analysis protocols — save model, diameter, bins and calibration to a file so a whole lab runs identical settings.
- Duplicate detection (SHA-256) so the same field is never counted twice.
CellCounter opens large libraries progressively, keeps previews and masks within a bounded working set, and moves expensive analysis away from the interface. Browsing and review remain responsive without loading every full-resolution image at once.
The complete feature list records the public functionality available in each released edition, including platform adaptations, preview features, and intentional limitations.
Requires macOS 15 or later. Universal binary (Apple silicon and Intel).
- Download
CellCounter-v1.0.13.zipfrom the macOS release. - Unzip and move
CellCounting.appinto Applications. (The application is called CellCounter; the bundle on disk is still namedCellCounting.app.) - The app is not notarized, so the first launch is blocked. Open System Settings → Privacy & Security, scroll to the bottom, and click Open Anyway. Full walkthrough: docs/INSTALL.md.
- Import images to inspect them immediately. When ready to detect cells, install a model from Models, return to Analysis setup, preview a representative image, then process the batch. Vendor formats need their compatible Python reader before import.
macOS says the app is "damaged"
This message normally means Gatekeeper has blocked the ad-hoc-signed, quarantined download. Confirm that the archive came from the official release, move the app to Applications, and run:
xattr -cr /Applications/CellCounting.appThen open it normally.
Requires 64-bit Windows 10 or Windows 11. Choose the .exe for a standard workstation installation or the
.msi for managed deployment. Both contain the same native application.
The installers are currently unsigned, so Windows SmartScreen may show an unknown-publisher warning. Verify the
download against the supplied SHA256SUMS.txt before continuing. The application provides three local model
families: Cellpose-SAM v2, Cellpose cyto3, and StarDist fluorescence. See the
Windows installation guide for model setup, upgrades, backup, and troubleshooting.
The web edition is a fully client-side PWA with no account, upload, telemetry, or application server. Download
the release archive and serve it from HTTPS or localhost, then install it from a current WebGPU-capable browser.
Its browser-native import, curation, measurement, assay, and export tools are available offline after installation.
Production model weights are not included in v0.1.0, so learned segmentation is intentionally unavailable in this preview. See the web guide for supported formats, browser requirements, and current limits.
For macOS v1.0.13:
- Open images with ⌘O, or a folder with ⌘⇧O. Choose a representative image in Analysis setup.
- Choose a task and model; confirm the pixel size and source channels. Install the model from Models if needed.
- Choose Preview this image to inspect masks, then Process batch. Use Import and inspect to defer detection.
- Open Processing to pause, resume, retry failures, or browse completed images. Review linked image/table/plot selections and correct masks.
- Compare saved mask variants, use Compare for two conditions, or export a PDF, CSV, ROI set or GeoJSON.
- Open Help → Keyboard Shortcuts (⌘/) for the available commands.
See the Windows guide and web guide for their respective workflows.
Native analysis runs on the computer where CellCounter is installed. The web edition processes imported images inside the browser. CellCounter has no image-upload or remote-inference path in these releases. Network access is used only when the operating system or user requests supporting software or model files; microscopy images are not included in those requests.
Important
CellCounter is a measurement tool; its built-in statistics are for exploration. Read this before a number from it goes into a paper.
The replication unit is the biological replicate, not the cell
The Compare tab's Mann-Whitney U test pools every individual cell across all images in a condition and treats them as independent. For condition-level inference this is pseudoreplication: it inflates n by orders of magnitude and returns very small p-values for biologically trivial differences.
For publication, aggregate first. Export the per-cell CSV, compute one summary per image (or per patient, or per well) — for example the median diameter — and test on those replicate-level values, or use a mixed-effects model with image or patient as a random effect. Treat the in-app pooled test as descriptive only.
No multiple-comparison correction
Comparing more than two conditions by re-selecting pairs gives uncorrected p-values and significance markers. Apply Holm or Benjamini-Hochberg (or an omnibus Kruskal-Wallis first) when reporting several contrasts.
Segmentation is not bit-for-bit reproducible across machines
Counts depend on the model version, the device (GPU or CPU), and the PyTorch and NumPy versions. Expect small run-to-run and machine-to-machine differences. For a reproducible methods section, record the model, the app version and the resolved dependency versions — the exported provenance sidecar captures model, calibration and parameters.
"Size" is an equivalent diameter
Each cell's size is the diameter of a circle with the same segmented area (2·√(area/π)) — a shape-agnostic
proxy, not a measured long or short axis. Per-cell "confidence" is a monotonic transform of Cellpose's
cell-probability, not a calibrated probability.
Assay results are withheld rather than guessed
Where a number cannot be produced honestly the app says so. For example, % marker-positive is withheld when the population shows no evidence of two distinct groups — an automatic threshold will otherwise split a single uniform population and report a confident, meaningless percentage.
Several tools share the name "Cell Counter". This project is not affiliated with, and is distinct from, the 2014 application CELLCOUNTER: Novel Open-Source Software for Counting Cell Migration and Invasion In Vitro (BioMed Research International, for Boyden-chamber assays) and the ImageJ / Fiji Cell Counter plugin (manual tally counting). CellCounter here is a Cellpose-driven counting and size-classification desktop application.
If CellCounter is useful in your work, please cite it (see CITATION.cff) and the segmentation model you ran:
- Stringer, C., Wang, T., Michaelos, M., & Pachitariu, M. (2021). Cellpose: a generalist algorithm for cellular segmentation. Nature Methods 18, 100–106.
- Pachitariu, M., & Stringer, C. (2022). Cellpose 2.0: how to train your own model. Nature Methods 19, 1634–1641.
- Using
cyto3 + restoreor Cellpose-SAM? Also cite the Cellpose 3 and Cellpose-SAM papers listed in the Cellpose repository. - Using StarDist or Omnipose? Cite their papers too.
Cellpose, StarDist, and Omnipose provide learned segmentation families. PyTorch, NumPy, SciPy, scikit-image, tifffile, and format-specific readers support the native analysis pipeline. Full inventory and licenses: THIRD_PARTY_LICENSES.md.
MIT — see LICENSE. Contributions welcome; see CONTRIBUTING.md.