Settings

The important settings are described on this page. See the Cellpose class for all run options.

Here is an example of calling the Cellpose class and running a list of images for reference:

from cellpose import models
from cellpose.io import imread

# model_type='cyto' or model_type='nuclei'
model = models.Cellpose(gpu=False, model_type='cyto')

files = ['img0.tif', 'img1.tif']
imgs = [imread(f) for f in files]
masks, flows, styles, diams = model.eval(imgs, diameter=None, channels=[0,0],
                                         flow_threshold=0.4, do_3D=False)

You can make lists of channels/diameter for each image, or set the same channels/diameter for all images as shown in the example above.

Channels

There are two channels inputs. The first channel is the channel you want to segment. The second channel is an optional channel that is helpful in models trained with images with a nucleus channel. See more details in the models page.

0=grayscale, 1=red, 2=green, 3=blue
0=None (will set to zero), 1=red, 2=green, 3=blue

Set channels to a list with each of these elements, e.g. channels = [0,0] if you want to segment cells in grayscale or for single channel images, or channels = [2,3] if you green cells with blue nuclei.

On the command line the above would be --chan 0 --chan2 0 or --chan 2 --chan2 3.

Diameter

The cellpose models have been trained on images which were rescaled to all have the same diameter (30 pixels in the case of the cyto model and 17 pixels in the case of the nuclei model). Therefore, cellpose needs a user-defined cell diameter (in pixels) as input, or to estimate the object size of an image-by-image basis.

The automated estimation of the diameter is a two-step process using the style vector from the network, a 64-dimensional summary of the input image. We trained a linear regression model to predict the size of objects from these style vectors on the training data. On a new image the procedure is as follows.

Run the image through the cellpose network and obtain the style vector. Predict the size using the linear regression model from the style vector.
Resize the image based on the predicted size and run cellpose again, and produce ROIs. Take the final estimated size as the median diameter of the predicted ROIs.

For automated estimation set diameter = None. However, if this estimate is incorrect please set the diameter by hand.

Changing the diameter will change the results that the algorithm outputs. When the diameter is set smaller than the true size then cellpose may over-split cells. Similarly, if the diameter is set too big then cellpose may over-merge cells.

Resample

The cellpose network is run on your rescaled image – where the rescaling factor is determined by the diameter you input (or determined automatically as above). For instance, if you have an image with 60 pixel diameter cells, the rescaling factor is 30./60. = 0.5. After determining the flows (dX, dY, cellprob), the model runs the dynamics. The dynamics can be run at the rescaled size (resample=False), or the dynamics can be run on the resampled, interpolated flows at the true image size (resample=True). resample=True will create smoother ROIs when the cells are large but will be slower in case; resample=False will find more ROIs when the cells are small but will be slower in this case. By default in versions >=1.0 resample=True.

The nuclear model in cellpose is trained on two-channel images, where the first channel is the channel to segment, and the second channel is always set to an array of zeros. Therefore set the first channel as 0=grayscale, 1=red, 2=green, 3=blue; and set the second channel to zero, e.g. channels = [0,0] if you want to segment nuclei in grayscale or for single channel images, or channels = [3,0] if you want to segment blue nuclei.

If the nuclear model isn’t working well, try the cytoplasmic model.

Flow threshold

Note there is nothing keeping the neural network from predicting horizontal and vertical flows that do not correspond to any real shapes at all. In practice, most predicted flows are consistent with real shapes, because the network was only trained on image flows that are consistent with real shapes, but sometimes when the network is uncertain it may output inconsistent flows. To check that the recovered shapes after the flow dynamics step are consistent with real ROIs, we recompute the flow gradients for these putative predicted ROIs, and compute the mean squared error between them and the flows predicted by the network.

The flow_threshold parameter is the maximum allowed error of the flows for each mask. The default is flow_threshold=0.4. Increase this threshold if cellpose is not returning as many ROIs as you’d expect. Similarly, decrease this threshold if cellpose is returning too many ill-shaped ROIs.

Mask threshold

The network predicts 3 outputs: flows in X, flows in Y, and cell “probability”. The predictions the network makes of the probability are the inputs to a sigmoid centered at zero (1 / (1 + e^-x)), so they vary from around -6 to +6. The pixels greater than the cellprob_threshold are used to run dynamics and determine ROIs. The default is cellprob_threshold=0.0. Decrease this threshold if cellpose is not returning as many ROIs as you’d expect. Similarly, increase this threshold if cellpose is returning too ROIs particularly from dim areas.

3D settings

Volumetric stacks do not always have the same sampling in XY as they do in Z. Therefore you can set an anisotropy parameter to allow for differences in sampling, e.g. set to 2.0 if Z is sampled half as dense as X or Y.

There may be additional differences in YZ and XZ slices that make them unable to be used for 3D segmentation. I’d recommend viewing the volume in those dimensions if the segmentation is failing. In those instances, you may want to turn off 3D segmentation (do_3D=False) and run instead with stitch_threshold>0. Cellpose will create ROIs in 2D on each XY slice and then stitch them across slices if the IoU between the mask on the current slice and the next slice is greater than or equal to the stitch_threshold.

3D segmentation ignores the flow_threshold because we did not find that it helped to filter out false positives in our test 3D cell volume. Instead, we found that setting min_size is a good way to remove false positives.