Cellpose 3

Overview

Cellpose 3 is the third major release of the Cellpose generalist cell segmentation framework, developed by Carsen Stringer and Marius Pachitariu at HHMI Janelia Research Campus and published in Nature Methods in 2025. The release delivers two complementary advances: a new super-generalist segmentation model called cyto3, trained jointly on nine diverse public datasets, and a family of one-click image restoration networks that denoise, deblur, and upsample microscopy images before segmentation.

The most distinctive contribution is the restoration network design philosophy. Rather than minimizing pixel-level reconstruction error — as in CARE and Noise2Void — Cellpose 3 trains its restoration networks to produce images that are optimally segmented by the Cellpose backbone. This shifts the optimization target from perceptual fidelity to downstream task performance, yielding restorations that consistently produce better cell masks even when they do not look the cleanest to the human eye.

Prior Cellpose versions were trained exclusively on the original Cellpose dataset. The cyto3 model extends this dramatically by training across nine public datasets spanning mammalian fluorescence imaging, yeast, and bacteria, making it substantially more robust out of the box for images that differ from standard cell culture microscopy.

Key Features

• cyto3 super-generalist model: Trained jointly across nine diverse public datasets covering mammalian cells (fluorescence and brightfield), yeast, and bacteria; substantially more robust than cyto2 on out-of-distribution images.
• Segmentation-optimized restoration loss: Compound loss combining a segmentation term (flow prediction error against the frozen Cellpose backbone) and a perceptual term (encoder activation correlation matrices), optimizing downstream mask quality rather than pixel reconstruction.
• One-click image restoration: Denoising, deblurring, and upsampling networks integrated directly into the Cellpose GUI and Python API, requiring no configuration or retraining by the user.
• No paired training data required: Restoration models are trained on synthetically degraded versions of clean public images, eliminating the need for matched noisy/clean experimental image pairs.
• High-throughput inference: The restoration network processes approximately 750 images per second on an A100 GPU, making it practical at dataset scale for large screening experiments.

Technical Details

The segmentation backbone is a U-Net with four downsampling and four upsampling blocks, each containing four convolutional layers with residual connections — identical in architecture to Cellpose 1 and 2. The cyto3 model uses this same architecture trained on nine public datasets: Cellpose cyto2 (~874 images), Cellpose nuclei (~1,136 images), TissueNet (~3,850 images), LiveCell (~4,704 images), Omnipose, YeaZ, and DeepBacs (~155 images), among others. Dataset sampling during training was weighted to prevent large homogeneous datasets from dominating gradient updates, with the original Cellpose cyto2 images sampled at approximately 59% probability.

The image restoration networks share the same U-Net design and are trained with a compound loss: a segmentation term measuring MSE between predicted and ground-truth spatial flows (scaled by a factor of five) plus binary cross-entropy on cell probability, combined with a perceptual term measuring deviations in encoder-level feature statistics from clean target images. Synthetic degradations used for training include Poisson noise at varying intensities, isotropic Gaussian blur (standard deviations of 1-10 pixels), and bilinear downsampling (factors of 2-7).

Quantitatively, cyto3 achieves an average precision at IoU 0.5 (AP@0.5) of 0.786 on the Cellpose test set versus 0.771 for a dataset-specific model. The denoising network outperforms Noise2Void and Noise2Self on held-out test sets, with segmentation performance approximately doubling on heavily degraded images. Validation on real experimental data — including fluorescence images of Drosophila wing epithelial cells and mouse cortical neurons — confirms that gains hold on biological images with non-synthetic noise.

Applications

Cellpose 3 is well suited for researchers working with noisy or degraded microscopy data. The denoising network recovers substantial segmentation performance from images acquired at low laser power or with high-sensitivity detectors. The upsampling network enables accurate segmentation of images acquired at lower magnification, while the deblurring network corrects Gaussian-profile blur common in widefield acquisitions. The cyto3 model covers mammalian cells (cultured and tissue-derived), yeast, and bacteria without retraining, making it a strong default choice for any light microscopy segmentation task. The combination of fast GPU inference and established Cellpose segmentation speed supports high-throughput imaging screens without workflow modifications.

Impact

Cellpose 3 extends one of the most widely adopted cell segmentation tools in biological imaging by addressing two persistent pain points: generalization to diverse cell types and robustness to image quality degradation. The segmentation-optimized restoration loss represents a conceptual departure from conventional image reconstruction that is likely to influence future biological image processing pipelines. One notable limitation is that the restoration networks are trained on Poisson noise, isotropic Gaussian blur, and uniform downsampling; non-isotropic blur, structured noise (stripe artifacts, autofluorescence), or sensor-specific patterns may not be corrected effectively. The cyto3 model also shows a small but consistent performance decrease (~0.015 AP@0.5) on the original Cellpose benchmark relative to a model trained exclusively on that dataset, a modest trade-off for broad generalization. Cellpose 3 is distributed through the standard Cellpose package (version 3.x) with no separate installation required.