Cellpose: a generalist algorithm for cellular segmentation

Carsen Stringer,Tim Wang,Michalis Michaelos,Marius Pachitariu

DOI: https://doi.org/10.1038/s41592-020-01018-x

IF: 48

2020-12-14

Nature Methods

Abstract:Many biological applications require the segmentation of cell bodies, membranes and nuclei from microscopy images. Deep learning has enabled great progress on this problem, but current methods are specialized for images that have large training datasets. Here we introduce a generalist, deep learning-based segmentation method called Cellpose, which can precisely segment cells from a wide range of image types and does not require model retraining or parameter adjustments. Cellpose was trained on a new dataset of highly varied images of cells, containing over 70,000 segmented objects. We also demonstrate a three-dimensional (3D) extension of Cellpose that reuses the two-dimensional (2D) model and does not require 3D-labeled data. To support community contributions to the training data, we developed software for manual labeling and for curation of the automated results. Periodically retraining the model on the community-contributed data will ensure that Cellpose improves constantly.

biochemical research methods

What problem does this paper attempt to address?

This paper attempts to solve several key problems in cell image segmentation:

1. **The universality problem of cell image segmentation**: Existing cell image - segmentation methods usually need to be trained for specific types of images, which limits their application on different datasets. The paper proposes a new general - purpose algorithm, Cellpose, aiming to accurately segment cells from multiple types of images without retraining the model or adjusting parameters for each type of image.

2. **Dealing with closely - arranged cells**: In many tissues, cells are closely arranged together and difficult to separate. Cellpose improves its ability to process such images through an improved neural network architecture and training strategy.

3. **Improving segmentation accuracy**: The paper introduces the design principles of Cellpose, especially how to generate a single smooth topological basin by simulating the diffusion process, and how to use the neural network to predict horizontal and vertical gradients to achieve more accurate cell contour tracking. These technological improvements significantly improve the segmentation accuracy, especially when dealing with cells with complex shapes.

4. **3D image segmentation**: The paper also explores the possibility of extending Cellpose to 3D image segmentation and proposes a method that does not require additional 3D - labeled data, using the existing 2D model to process 3D images.

5. **Community contribution and continuous improvement**: In order to further improve the performance of Cellpose, the paper encourages users to contribute manually - labeled image data to regularly retrain the model, ensuring that it can continuously adapt to new image types and challenges.

Overall, the main goal of this paper is to develop a general - purpose algorithm that can be widely used in different types of cell image segmentation, while providing a high - performance and easy - to - use solution.