Introduction to Cellpose

Introduction to Cellpose#

In this section, we dive into Cellpose, a state-of-the-art deep learning tool for biological image segmentation. You’ll learn how it works, what makes it unique compared to other models like U-Net and StarDist, and how to use it via its easy-to-use web interface (GUI). We’ll conclude by setting the stage for using Cellpose in Google Colab for more flexible and scalable workflows.

What is Cellpose?#

Cellpose is a deep learning model for segmenting biological images. It was released in 2021 by Stringer et al. with a specific goal: to create a model that works well on many types of microscopy data without needing retraining.

Unlike tools like U-Net or StarDist that often require dataset-specific training, Cellpose was designed as a generalist model. It was trained on a large, diverse set of images—fluorescent images, phase contrast, cells with or without nuclei, cytoplasmic markers, and more.

The Core Innovation: Flow Fields#

Key Idea#

Cellpose predicts spatial flow fields instead of binary masks. Each pixel in the image is assigned a vector that “points” toward the center of its object. After these flows are computed, the pixels are grouped based on how they flow together—like tracking water to a drain.

How it Works (Conceptually)#

Input: An image of cells, nuclei, or tissue.
Model predicts:
- A probability map: which pixels are likely to belong to any object.
- A flow field: where each pixel would move if it were pulled toward its object’s center.
Pixels are grouped: Based on where they flow, pixels are grouped into instances (cells).

Cellpose Training Data and Generalization#

Cellpose was trained on over 70,000 manually annotated cells across various imaging modalities, including:

Fluorescence (cytoplasmic, nuclear, and membrane markers)
Brightfield
Phase contrast
Tissue sections

This diversity is what makes Cellpose a generalist—you often don’t need to retrain it on your own data.

Key strength, Cellpose allows custom training if needed, which is something we’ll briefly touch on later.

Cellpose Timeline#

Version	Release Date	Name / Paper	Key Features & Innovations	Use Cases / Impact
Cellpose 1.0	December 2020	[Stringer et al., Nat. Methods (2020)]	- Introduced flow-based segmentation - Generalist model trained on a wide range of image types - Pretrained on nuclei and cytoplasm - GUI and command-line interface (CLI) support	- Ready-to-use for diverse cell types - No need for custom training - Reproducible and scalable
Cellpose 2.0	April 2022	[Pachitariu & Stringer, Nat. Methods (2022)]	- Enabled human-in-the-loop training - Easy custom model training without code - Introduced model zoo for diverse pretrained options - Improved GUI with model selection and training support	- Great for labs with specific cell types - Enables personalized models for better accuracy - More flexible for experimental variation
Cellpose 3.0	February 2024	[Stringer & Pachitariu, Nat. Methods (2025)]	- Added image restoration before segmentation - Enhanced segmentation quality for noisy/low-contrast data - Integration into GUI/API/CLI - Preprocessing saved as `_seg.npy`	- Particularly effective for challenging or degraded images - One-click enhancement for better masks
Cellpose-SAM	May 2025	[Pachitariu, Rariden & Stringer, bioRxiv (2025)]	- Achieves superhuman generalization - Works on 2D and 3D data - Handles noise, blur, contrast inversion, variable scales - Compatible with Hugging Face online batch processing	- Broadest generalization across datasets - Ideal for complex, noisy, real-world images - Scalable deployment with cloud integration

Using Cellpose via the GUI#

Cellpose comes with a graphical user interface (GUI) that lets you segment images without any coding. It’s one of the easiest ways to get started.

Accessing the GUI#

You can access the Cellpose GUI in two main ways:

Option 1: Online Web Tool Cellpose developers provide a web-based GUI:

Go to https://www.cellpose.org
Click “Run Cellpose GUI”. This launches a web app via the browser.

Note: The web app uses limited resources and is best for trying out single images.

Option 2: Local Install (Desktop App) If you want better performance:

Install Cellpose on your local computer
Launch the GUI using cellpose from the command line

Instructions for installation are available on MouseLand/cellpose