Segmenting densly packed cells with StarDist

Another versatile and powerful segmentation tool is StarDistꜛ. StarDist is based on a combined method of neural networks and probability density estimation. It utilizes a neural network to identify the positions of cell nuclei and perform an initial segmentation. Subsequently, a probability density estimation is applied to the initial segmentation to generate precise cell contours. StarDist is particularly useful for segmenting heavily overlapping or densely packed cells. In this tutorial, we will use StarDist for segmenting fine cell structures such as nuclei.

Using StarDist in Napari

StarDist comes with the stardist-napariꜛ Napari plugin. Its usage is as straightforward as for any other Napari plugin. Simply open the StarDist plugin widget via the Plugins menu, select the input image layer, select the appropriate model for your underlying image (2D or 3D) and specify the axes of your image. When you are done, just the Run button. The run button is grayed out when one or more properties are not set correctly.

Applying StarDist to the sample image “test_lsm_ch0” provided in the GitHub data folder (see image credits in the acknowledgements section below).

The segmentation result will be displayed as new layers in the layer list and you can evaluate the segmented image parts with the Measurement function from the Tools->Measurement tables menu (see “Feature extraction” in Chapter “Image segmentation and feature extraction”).

Comparing StarDist with Cellpose

While Cellpose is also able to segment cell nuclei, for the underlying image, StarDist achieves a better segmentation result:

Comparing StarDist to Cellpose. Top right: Original image. Top left: StarDist segmentation. Bottom left: Cellpose segmentation (cytoplasm). Bottom right: Cellpose segmentation (nuclei).

This is due to the fact that StarDist is trained on a larger and more diverse data set. However, StarDist is not able to segment the cytoplasm of the cells. In general, the segmentation performance also dependents on the underlying image complexity, thus, one should not commit to only one method, but try different methods and compare the results.

Further readings

• GitHub repository of the Napari StarDist pluginꜛ including a detailed documentation
• GitHub repository of StarDistꜛ
• StarDist’s Napari hub pageꜛ
• StarDist tutorial on multi-class classificationꜛ
• Uwe Schmidt, Martin Weigert, Coleman Broaddus, and Gene Myers. Cell Detection with Star-convex Polygons. International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), Granada, Spain, September 2018.
• Martin Weigert, Uwe Schmidt, Robert Haase, Ko Sugawara, and Gene Myers. Star-convex Polyhedra for 3D Object Detection and Segmentation in Microscopy. The IEEE Winter Conference on Applications of Computer Vision (WACV), Snowmass Village, Colorado, March 2020

Acknowledgements

The data file used in this tutorial is an excerpt from the example data set of Napari PartSeg pluginꜛ