henhaobio
Sign in

Nature MethodsBioinformaticsPeer-reviewed journal article

Cellpose: a generalist algorithm for cellular segmentation

Carsen StringeriD, Tim Wang, Michalis MichaelosiD, Marius PachitariuiD

PDFDOIPreprintCode
In-depth full-text digest available · Jump to digest ↓

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. We trained Cellpose on a new dataset of highly-varied images of cells, containing over 70,000 segmented objects. We also demonstrate a 3D extension of Cellpose which reuses the 2D model and does not require 3D-labelled data. To support community contributions to the training data, we developed software for manual labelling and for curation of the automated results, with optional direct upload to our data repository. Periodically retraining the model on the community-contributed data will ensure that Cellpose improves constantly.

In-depth summary

Claude Opus 5.5 · xhigh effort · 5 Nov 2026 · from the full text of version 2 (1 Apr 2020) · CC BY-NC 4.0

Cellpose segments diverse 2D and 3D microscopy cell images without retraining by predicting topological gradient vector flows via a style-conditioned residual U-net and routing pixels to cell centers.

Deep learning models for cellular segmentation typically require specialized training data for each microscopy modality and struggle with complex, non-convex cell shapes. By converting ground-truth masks into simulated heat-diffusion vector flows and training a style-conditioned residual U-net on over 70,000 segmented objects across diverse image types, Cellpose generalizes out-of-the-box to 2D and 3D cellular images without user retraining.

Key results

  1. ShownCellpose outperforms StarDist and Mask R-CNN when trained and evaluated as a specialist model on morphologically complex neurons.

    On the 100-image CellImageLibrary dataset of two-channel cultured neurons, specialist Cellpose matched 485 of 521 ground-truth regions of interest with 18 false positives at an IoU threshold of 0.5, reaching an average precision (AP) of 0.88 versus 0.76 (StarDist) and 0.80 (Mask R-CNN). · Training dataset; Benchmarks; Specialized data; Figure legends

  2. ShownA single generalist Cellpose model substantially outperforms generalist StarDist and Mask R-CNN across diverse microscopy and non-microscopy images while matching specialist accuracy.

    At IoU = 0.5, the generalist Cellpose model achieved an AP of 0.82 across all test cell images versus 0.68 for StarDist and 0.70 for Mask R-CNN, an AP of 0.79 vs 0.67 (StarDist) and 0.66 (Mask R-CNN) on generalized cell images excluding CellImageLibrary, and an AP of 0.73 vs 0.54 and 0.55 on visually distinct non-microscopy and atypical microscopy images. · Benchmarks; Figure legends

  3. ShownTwo-step calibration from learned neural network image style vectors accurately predicts median object diameters on unseen test images.

    Combining linear regression on the 256-dimensional global average pooled style vector with a first-pass Cellpose segmentation achieved correlations of r = 0.93 on the cytoplasmic test set (69 images) and r = 0.97 on the nuclear test set (111 images) with ground-truth object diameters. · Nucleus data; Deep neural network; Test time resizing; Figure legends

  4. ShownAveraging 2D flow predictions across orthogonalXY, XZ, and YZ slices extends Cellpose to 3D volumetric segmentation without 3D-labelled training data.

    On a ~190 x 190 x 190 um expansion-microscopy volume of mouse cortical neurons containing 183 manually annotated cells, 3D Cellpose predicted 172 cells with 17 false positives at IoU = 0.5 and outperformed a manually optimized 3D ilastik pipeline across all IoU thresholds. · 3D segmentation; 3D data; 3D model; Figure legends

Methods, data and reproducibility

Converted ground-truth masks via simulated heat diffusion into horizontal and vertical topological gradients. Trained a 4-scale residual U-net in MXNet with a 256-D style vector (500 epochs, SGD) to predict horizontal/vertical flows and cell probabilities, using two-step size calibration, 4-model ensembling, 224x224 tiling, flow-consistency filtering (0.4), and 200-step flow tracking in 2D or XY/XZ/YZ 3D.

  • Generalist dataset of cytoplasmic fluorescence (n=316), brightfield (n=50, OMERO), membrane (n=58, Micro-Net), other microscopy (n=86), and non-microscopy images (n=98) · multiple species · n = 616 images (>70,000 segmented objects; 69 reserved for testing)
  • Assembled nuclear segmentation dataset from Data Science Bowl BBBC038v1, BBBC039v1 (U2OS), MoNuSeg (human H&E organs), and ISBI 2009 (U2OS and NIH3T3) · multiple species · n = 1,139 images (111 reserved for testing)
  • Two-fold expanded mouse cortical neuron 3D tissue volume co-stained with DAPI and 28S rRNA fluorescent oligonucleotide probe · mouse · n = 1 cube (~190 x 190 x 190 um, 183 manually segmented cells)
Methods & tools
Cellpose, MXNet, StarDist, Mask R-CNN, ilastik, OpenCV, NumPy, SciPy, Numba, PyQt, pyqtgraph, scikit-image
Software & environment
Python 3, numpy, scipy, numba, opencv, mxnet, PyQt, pyqtgraph, scikit-image, matplotlib, jupyter-notebook
Reproducibility gaps
The manually segmented cytoplasmic dataset is stated to be made available upon publication; the web server at cellpose.org omits test-time tiling and augmentations to reduce compute time.
Limitations (2)
  • Cellpose is not designed for images containing objects with highly heterogeneous shapes and textures within the same field of view, or where occlusions require overlapping masks since each pixel is assigned to at most one mask. · stated by the authors
  • Segmentation accuracy is lower on visually distinct non-microscopy and atypical microscopy images (AP 0.73 at IoU = 0.5) than on canonical fluorescent cell datasets. · stated by the authors

Read this if you analyze 2D or 3D biological microscopy images and need a generalist cell and nucleus segmentation tool that works without retraining.

Citations
4,371
FWCI
138.79
Field percentile
top 1%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISPachitariu '22Stringer '25Israel '23Dolev '19Xun '22David '26
80% semantic matchBioinformatics
Marius Pachitariu, Carsen Stringer·Nature Methods·Nov 7, 2022· 1,400 citations