Part 5: What about deep learning

Part 5: What about deep learning#

Deep learning-based segmentation is getting extremely good, so why learn “the classics”? A few reasons:

  1. It’s important to know the basics to be able to troubleshoot when advanced methods fail.

  2. The fundamentals make it easier to understand how these kinds of data are represented. Segmentation approaches all start and end with the same representation. PyTorch arrays are basically like NumPy arrays, with a few extra bells and whistles.

To see what we mean, let’s try out a state of the art tool for segmentation of cellular images.

Exercise: cellpose (2D)#

Warning

To run this exercise, you need to install an additional dependency either via

conda install cellpose

or

pip install cellpose

depending on your chosen environment.

Cellpose is a deep-learning based segmentation tool for images of cells. It’s outstanding at segmenting 2D images, so you should try it on 2D datasets from scikit-image:

https://scikit-image.org/docs/dev/auto_examples/data/plot_scientific.html

You can also grab a slice of the 3D cells (which would also be a nice refresher for chapter 0!).

As input, cellpose requires a 2D image with channels as the final axis.

  • Read the Cellpose documentation example notebook here.

  • (Optional) index the middle plane of your image if it’s 3D.

  • (Optional) Use np.transpose to ensure the channel axis to the final position in the array.

  • Call cellpose with the appropriate channel settings (carefully read the example in the Cellpose documentation).

  • Display the image channels and the segmentation masks overlaid in napari.

# solution ...
Contents