Once the script has been loaded into the operator we can set up the parameters for the operation.

All the parameters here are defined by the cellpose method. More information about each of those settings can be found on the cellpose website.

As mentioned in the introduction, cellpose uses pre-trained models to segment the cells. Some pre-trained models are included (cyto2/nuclei), others can be downloaded from the cellpose website. For most applications, the cyto2 or nuclei models work fine. To select either of these models, we can simply type the name of the model in the Model_Name field (note that the name is case sensitive).

As hinted at by the name of these models, they are each tuned to detect specific portions of cells. Namely, the cyto2 model detects cytoplasm, while the nuclei model is tuned for nuclei. The Channel field should therefore indicate which channel shows this specific component. Note that the cyto2 model can also produce improved results by adding a secondary channel for the nuclei. In this case, the cytoplasm channel should be selected in the Input_channel field, while the nuclear channel should be used in the Second_channel field. If there is no nucleus channel, or the model used is not able to use a second channel the Second_channel field should be set to 0 (zero).

Note that the channel numbering is taken from their ordering in the set. This can be clearly seen in the Channel Visibility panel.

In this case, the DAPI nuclear channel is channel #1.

The Diameter_in_um field is used to define how big we expect the objects to be. This does not have to be an exact specific number, but it should be representative of the typical object diameter. Some experimentation with this value is worthwhile to better understand how much of an effect it has. Generally, the best way to set this is to use the Measure tool to measure the diameter of a typical cell.

The Flow_threshold and Cellprob_threshold fields are used to fine-tune the predictions for our specific images. Details of how each of these parameters affects the result can be found here, but as with the diameter, some experimentation is worthwhile. Note that the Cellprob_threshold value must be between -6 & +6.

The Fast_3D parameter is binary and can therefore only be on or off. If it is left on the predictions are made plane by plane, without any knowledge of the neighboring planes, and the end result is simply collating those predictions in 3D. As the name implies this is faster but potentially less accurate as an error in the prediction in one plane is more likely to lead to an object being split in the Z axis. In contrast, if this option is left off the model uses information from neighboring planes to improve the prediction in the current plane. In either case, the end result will be a 3D object (if the set has only one plane the object is simply one plane thick). 

The operator can be configured to use custom models. If we want to use a custom model we can simply check the box for Custom_model and enter the path for the model file (note path names must use forward slashes e.g. "C:/CustomPath/Model"). The name in the Model_name field will then be ignored. More information on creating and using custom models can be found on the cellpose website.

Finally, we have two Vision4D-specific parameters. The first is the Tile_Size. This implementation of cellpose has been designed to take advantage of Vision4D's capacity to process very large images. However, the cellpose model itself is not built to be RAM independent the way that Vision4D is. Therefore, to process very large images this implementation divides the images into multiple tiles, each tile is processed individually and the results are collated from these predictions. Because these models are RAM-dependent, the tile size should be set according to the available memory. In practice, leaving the tile size at the default value works for most people.

The final parameter is common to all segmentation operations, it is the tag used to label the objects created by this operation. It is how other pipeline operations can call on the objects created by this operation to use or refine the results.