Cell counting is one of the most fundamental tasks in any biology lab. Whether you're passaging cells, setting up assays, or analysing drug responses, you need to know how many cells you have and how healthy they are. But the methods have evolved dramatically — from manual haemocytometers to sophisticated AI-powered microscopy systems.
In this guide, we explore the full spectrum of cell counting technologies: from classic watershed algorithms in ImageJ, through modern deep learning approaches like CellSAM and Cellpose-SAM, to fully integrated systems like the EVOS M3000 that count cells and measure confluence automatically with a single tap.
🔬 The Classic Approach: Watershed Algorithms
For decades, the go-to method for automated cell counting in microscopy has been the watershed algorithm. Originally developed for topographic analysis, watershed segmentation treats fluorescence intensity as "elevation" — bright cells are peaks, dark background is valleys. The algorithm "floods" the image from local minima and places boundaries where different flooding regions meet.
How Watershed Cell Counting Works
- Step 1 — Preprocessing: Gaussian blur to reduce noise, background subtraction to normalise illumination
- Step 2 — Thresholding: Convert grayscale to binary mask (cells = white, background = black)
- Step 3 — Distance Transform: Calculate distance from each pixel to the nearest background pixel. Cell centres become local maxima
- Step 4 — Watershed: Treat distance transform as topography and flood from maxima. Boundaries form where different catchment basins meet
- Step 5 — Counting: Each segmented region = one cell
ImageJ & Fiji Watershed Implementation
The Find Maxima and Watershed tools in ImageJ/Fiji have been the workhorse for academic labs. The built-in Cell Counter plugin allows semi-automated counting with manual verification, while the TrackMate plugin extends this to time-lapse sequences.
For automated workflows, the Analyze Particles function combined with Binary → Watershed can segment touching cells. However, this requires careful parameter tuning for each cell type — what works for HEK293 cells may fail completely with primary neurons or tissue sections.
Limitations: Watershed struggles with densely packed cells, irregular morphologies, and overlapping nuclei. It also requires consistent staining and imaging conditions — real-world lab variability often breaks the algorithm.
Get it: ImageJ/Fiji (Free Download) — Cell Counter Plugin — Watershed Guide
🤖 The AI Revolution: Deep Learning Cell Segmentation
Modern deep learning approaches have transformed cell counting from a parameter-tuning nightmare into a robust, generalisable workflow. These models learn features directly from training data rather than relying on handcrafted rules.
Cellpose: The Generalist Cell Segmentation Model
Cellpose (Stringer et al., 2021) was a breakthrough — a generalist algorithm that segments cells "out-of-the-box" across diverse microscopy images without manual parameter tuning. The latest iteration, Cellpose 3.0, adds one-click image restoration (denoising and deblurring) to improve segmentation quality on degraded images.
- Architecture: Custom neural network trained on a broad range of cell types and imaging modalities
- Generalisation: Works on cells it was never trained on — truly zero-shot performance
- Speed: GPU-accelerated, processes images in seconds
- Open Source: Python package with napari integration for interactive visualisation — Cellpose.org (Free)
In 2025, Cellpose-SAM was introduced — combining Cellpose with Meta AI's Segment Anything Model (SAM) for "superhuman generalisation" that exceeds inter-human annotation agreement in quality.
CellSAM: A Foundation Model for Cell Segmentation
Published in Nature Methods (December 2025), CellSAM represents the next generation of foundation models for biological imaging. Unlike task-specific models, foundation models are trained on massive, diverse datasets and can be adapted to new tasks with minimal fine-tuning.
CellSAM Key Features
- Foundation Model Architecture: Pre-trained on millions of diverse cell images
- Broad Generalisation: Handles cell types and imaging conditions not seen during training
- Minimal Annotation: Requires only a few examples to adapt to new cell types
- Multi-Modal Input: Processes fluorescence, brightfield, and phase contrast images
- Open Weights: Available for researchers to fine-tune on their specific data
Publication: Marks et al., "CellSAM: a foundation model for cell segmentation", Nature Methods, 2025. Read Paper →
Get it: CellSAM GitHub (Free)
Meta AI Segment Anything for Microscopy (μSAM)
μSAM (micro-SAM) builds on Meta AI's Segment Anything Model (SAM) — a general-purpose image segmentation model released in 2023. μSAM specialises SAM for microscopy and biomedical imaging through fine-tuning on cellular datasets.
- Interactive Segmentation: Click a cell and μSAM segments it automatically
- 2D and 3D Support: Handles both standard microscopy and volumetric datasets
- napari Integration: Works within the popular napari image viewer for interactive workflows
- Automatic Annotation: Rapidly creates training data for custom models
Publication: "Segment Anything for Microscopy", Nature Methods, 2024. Read Paper → — Get μSAM (Free)
SAMCell: Label-Free Cell Segmentation
SAMCell (2025) demonstrates that SAM can segment cells even without fluorescent labels — using only brightfield or phase contrast images. This opens cell counting to label-free workflows where staining is impossible or undesirable.
Publication: "SAMCell: Generalized label-free biological cell segmentation with segment anything", PLOS One, 2025. Read Paper →
⚡ Integrated Solutions: EVOS Automated Cell Counting
While open-source AI tools offer incredible flexibility, many UK labs need turnkey solutions that work out-of-the-box without Python scripting or GPU setup. The EVOS M3000 Imaging System integrates automated cell counting and confluence detection directly into the microscope.
EVOS M3000 Cell Counting Features
- One-Tap Counting: Touch the screen and the system counts cells automatically — no parameters to tune
- Live/Dead Discrimination: Distinguishes viable cells from dead cells using fluorescence viability stains
- Confluence Detection: Measures the percentage of culture vessel covered by cells in real time — under 1 second
- Size and Morphology: Reports cell diameter, area, and roundness for quality control
- No External PC: All analysis runs on the integrated touchscreen — no software installation needed
Why Integrated Systems Matter
For busy cell culture facilities, the workflow integration is as important as the algorithm. The EVOS M3000:
- Eliminates Data Transfer: Images are captured, analysed, and reported on the same device
- Standardises Protocols: Every technician gets identical results — no subjective manual counting
- Tracks Over Time: Historical data shows growth curves and passaging schedules
- Validates Experiments: Normalise assays by actual cell number, not just passage number
Real-World Impact
A typical UK research lab running 50 cell culture flasks per week might spend 5–10 hours on manual counting using a haemocytometer. With automated EVOS counting, this drops to under 30 minutes — freeing researchers for actual science while improving reproducibility.
Confluence detection is equally transformative. Instead of subjective "70% confluent" estimates that vary between technicians, the EVOS reports exact percentages — enabling precise passaging schedules and consistent experimental timing.
📊 Algorithm Comparison: Which Method for Your Lab?
| Method | Best For | Expertise | Cost | Accuracy |
|---|---|---|---|---|
| Manual Haemocytometer | Quick checks, teaching labs | None | £ | Variable (±20%) |
| ImageJ Watershed | Research labs with standard cell lines | Moderate | Free | Good (±10%) |
| Cellpose 3.0 | Diverse cell types, custom workflows | Moderate | Free (GPU needed) | Excellent (±5%) |
| CellSAM | Novel cell types, minimal annotation | Advanced | Free (GPU needed) | Superhuman |
| μSAM | Interactive segmentation, 3D data | Moderate | Free (GPU needed) | Excellent |
| EVOS M3000 | Busy labs, standardisation, compliance | None | ££ (system cost) | Excellent (±5%) |
🧬 Emerging Methods: What's Next in 2026
Stamped Counting for Dense Cell Populations
Stamped counting (2025) uses a novel approach where cells are "stamped" onto a regular grid for density estimation — ideal for extremely dense cultures where individual segmentation fails. Published in Methods in Microbiology, 2025.
Dynamic Label Assignment (DLA-Count)
DLA-Count (2025) addresses dense cell distributions where overlapping nuclei make traditional methods fail. Uses dynamic label assignment networks to resolve touching cells. arXiv Paper →
Two-Step Deep Learning (DeepCellCount)
DeepCellCount (2025) combines detection and regression in a two-step pipeline — first localising cells, then estimating counts. Particularly effective for heterogeneous populations. Read Paper →
🦐 The Bottom Line
Cell counting has evolved from manual counting chambers → classical algorithms → deep learning → integrated hardware. The best choice depends on your lab's needs:
- Teaching or occasional use: Manual haemocytometer or ImageJ watershed
- Research with diverse samples: Cellpose 3.0 or CellSAM on your workstation
- High-throughput screening: Cellpose-SAM or μSAM for batch processing
- Busy cell culture facility: EVOS M3000 for turnkey automation and compliance
The trend is clear: AI is making cell counting more accurate, more generalisable, and more accessible. Whether you choose open-source tools or integrated systems, the days of manual counting are numbered — and that's good news for reproducibility.
Not sure which approach fits your workflow? The Shrimp & Scan team evaluates cell counting solutions for UK labs — from free ImageJ plugins to integrated EVOS systems. Drop us a message and we'll help you choose.
📚 Further Reading & Resources
- ImageJ Cell Counter Plugin Documentation
- ImageJ Watershed Separation Guide
- Cellpose 3.0 Official Website
- CellSAM: Nature Methods Paper (2025)
- Cellpose-SAM: bioRxiv Preprint (2025)
- μSAM: Segment Anything for Microscopy
- μSAM: Nature Methods Paper (2024)
- SAMCell: PLOS One Paper (2025)
- Thermo Fisher: EVOS M3000 Confluency Tool Video
- Thermo Fisher: EVOS M3000 Product Page